Next-Generation 100-Gigabit Metro Ethernet (100 GbME ... - CiteSeerX

12 downloads 189510 Views 1MB Size Report
of Ethernet to metropolitan-area distances. Such an implemen- tation of Ethernet in the core may yield cost benefits through a less complex network design with ...
2420

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Next-Generation 100-Gigabit Metro Ethernet (100 GbME) Using Multiwavelength Optical Rings Alejandra Zapata, Student Member, IEEE, Michael Düser, Member, IEEE, Jason Spencer, Student Member, IEEE, Polina Bayvel, Senior Member, IEEE, Ignacio de Miguel, Member, IEEE, Dirk Breuer, Norbert Hanik, Member, IEEE, and Andreas Gladisch, Member, IEEE

Abstract—This paper investigates the challenges for developing the current local area network (LAN)-based Ethernet protocol into a technology for future network architectures that is capable of satisfying dynamic traffic demands with hard service guarantees using high-bit-rate channels (80. . .100 Gb/s). The objective is to combine high-speed optical transmission and physical interfaces (PHY) with a medium access control (MAC) protocol, designed to meet the service guarantees in future metropolitan-area networks (MANs). Ethernet is an ideal candidate for the extension into the MAN as it allows seamless compatibility with the majority of existing LANs. The proposed extension of the MAC protocol focuses on backward compatibility as well as on the exploitation of the wavelength domain for routing of variable traffic demands. The high bit rates envisaged will easily exhaust the capacity of a band and will require network algosingle optical fiber in the rithms optimizing the reuse of wavelength resources. To investigate this, four different static and dynamic optical architectures were studied that potentially offer advantages over current link-based designs. Both analytical and numerical modeling techniques were applied to quantify and compare the network performance for all architectures in terms of achievable throughput, delay, and the number of required wavelengths and to investigate the impact of nonuniform traffic demands. The results show that significant resource savings can be achieved by using end-to-end dynamic lightpath allocation, but at the expense of high delay. Index Terms—Dynamic optical networks, Ethernet, metropolitan-area networks (MANs), optical burst switching (OBS), ring networks, wavelength-routed optical networks.

Manuscript received December 15, 2003; revised July 7, 2004. This work was supported by Deutsche Telekom Innovations Management within the framework of the project ONW2001+. This work was also supported in part by a Presidente de la Republica Scholarship (Chile), the Ian Karten Charitable Trust (United Kingdom), and Marconi Corporation plc (Coventry, U.K.) A. Zapata, M. Düser, J. Spencer, and P. Bayvel are with the Department of Electronic and Electrical Engineering, University College London (UCL), London WC1E 7JE, U.K. (e-mail: [email protected]; [email protected]; [email protected]; [email protected]). I. de Miguel is with the Departamento de Teoría de la Señal y Comunicaciones e Ingeniería Telemática, Universidad de Valladolid 47011 Valladolid, Spain (e-mail: [email protected]). D. Breuer and A. Gladisch are with the Technologiezentrum, T-Systems Nova GmbH, Deutsche Telekom, Berlin, Germany (e-mail: [email protected]; [email protected]). N. Hanik was with Technologiezentrum, T-Systems Nova GmbH, Deutsche Telekom, Berlin, Germany. He is now with Lehrstuhl für Nachrichtentechnik, Fachgebiet Leitungsgebundene Übertragungstechnik, Munich University of Technology, 80290 Munich, Germany (e-mail: [email protected]). Digital Object Identifier 10.1109/JLT.2004.836809

I. INTRODUCTION

F

ROM the first proposal in 1973, Ethernet has evolved very rapidly from providing short-distance connections between computers in local area networks (LANs) to cover campus-size distances and beyond in the 10-Gigabit Ethernet (10 GbE) standard (IEEE Standard 802.3ae). The main advantages of Ethernet are low cost, simplicity, high speed, and strong market penetration compared with other protocols such as Fiber Distributed Data Interface (FDDI) [1]. However, today, the connection beyond buildings and campuses is still performed over a combination of asynchronous transfer mode (ATM) and synchronous digital hierarchy (SDH) networks, usually by transporting layer 3 protocol data units, such as Internet protocol (IP) packets, rather than actual Ethernet frames. This will change, however, with the advent of 10 GbE, which allows point-to-point connections of up to 40 km, extending the reach of Ethernet to metropolitan-area distances. Such an implementation of Ethernet in the core may yield cost benefits through a less complex network design with fewer layers and provide a simpler, more scalable, homogenous campus-to-campus and metropolitan-area network (MAN) [2]. Inevitably, there are problems with the current versions of Ethernet as a protocol for the MAN. The legacy of Ethernet as a LAN protocol still prevails in the latest version of Ethernet (IEEE 802.3ae [3]), which does not support MAN functionality (e.g., resource allocation and advanced routing) and is still based on limited addressing and a maximum frame size of 1500 B, possibly too small for high-speed applications. This paper investigates the key requirements to extend Ethernet networks to 100-Gigabit Metro Ethernet (100 GbME), in terms of scalability and functionality, beyond the existing metro Ethernet initiatives, such as the evolving resilient packet ring (RPR) standard [4]. The proposed 100 GbME for future MANs will be based on optical-ring architectures supporting new dynamic optical-routing transport protocols. Initially, the ring topologies are the preferred architecture by the network operators for implementing MAN networks because they are easier to deploy and manage than arbitrarily meshed network architectures. This paper shows that the next generation of Ethernet with a tenfold increase in channel speed of 100 Gb/s will require fundamentally different network design, taking into account increased bandwidth consumption and incorporating network control and

0733-8724/04$20.00 © 2004 IEEE

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

management into the high-capacity MAN protocol. The essential elements of a medium-access control (MAC) protocol, required to control and manage these 100-GbME networks, are discussed in Section II. To achieve the required resource optimization and reuse, dynamic optical networking will be required, and four different architectures to fulfill these functions are proposed in Section III. Their performance is evaluated and compared in Section IV. The physical implementation aspects related to 100GbME networks and physical interfaces (PHY) are discussed in Section V. The results of this study will not only be of interest for the design of 100 GbME but also yield valuable insights into the design of high-speed networks over moderate metro distances. It is likely that 100 GbME will be implemented first in the MANs, although it may subsequently be extended to long-haul distances, once technology is sufficiently mature to allow this. II. ETHERNET OVERVIEW AND EXTENSIONS Ethernet is a standard that defines a connectionless, routable, variable-packet-size data-link layer protocol and a series of physical-layer interfaces. It was originally developed to connect hosts in a LAN over short distances ( 100 m) over physical layers with a shared-bus architecture that had no loops in its topology and no central control system. As a connectionless protocol, it offers no guarantees as to packet loss, delay, or any other end-to-end characteristics. In more recent implementations, Ethernet has moved from contented shared-bus cabling to hub-based unshielded cabling and hub-based fiber. The shift to a hub-based architecture has meant that the actions of a hub now determine network capability and performance, as all contention over resources now lies within the hub. During this evolution, faster switching (in the electrical domain) and higher bandwidth physical links (converging on optical-only physical layers) have seen Ethernet move out from the LAN applications into the campus where technologies such as FDDI were typically used [1]. Ethernet application in larger and more complex LANs has, however, identified limitations in the original design and required extensions to the original protocol to support new features. The introduction of filtering in hubs prevented packets from flooding every hub interface and allowed arbitrary physical topologies. The use of a spanning tree algorithm (IEEE Standard 802.1D [5]) removed loops in the logical topology and, therefore, always formed a tree with a single shared route to each destination. Standard 802.1D also added the ability to monitor links and passively learn a table of accessible Ethernet hosts. Later, additional Ethernet switch (a filtering hub) features were introduced, such as the ability to aggregate multiple physical links between switches (IEEE Standard 802.3ad [6]) to aid the scaling of capacity as well as traffic control extensions such as traffic prioritization (IEEE Standard 802.1p [7]) and flow control between switches (IEEE Standard 802.3x [8]). These extensions have all defined changes to switch behavior, but in IEEE Standard 802.3ac [9], a new frame format added support for multiple Ethernet networks over the same physical links through the use of virtual LAN tagging. In a virtual LAN (VLAN), the Ethernet stations can be seen as a series of overlaid Ethernet topologies, each with their own routing and logical topologies (IEEE Standard 802.1s [10]).

2421

In contrast to LANs, metro networks impose fundamentally different demands, as discussed in [11]. They are expected to carry more than just best-effort data, as well as to support a very large number of end users and high-capacity links in a scalable manner. In addition, transport network features are expected, such as operations, administration, and maintenance (OAM) functionality, improved network resilience, and quality-of-service (QoS) guarantees. The main new features required of metro Ethernet are described hereafter. Scalability: The nonhierarchical (nonsubnetworked) nature of MAC address (an Ethernet address) space does not allow for route aggregation in switches, and therefore, every switch must contain an entry for every Ethernet node, which can lead to very large tables (routing table scales linearly with network size). This can, however, become less significant by using layer 3 routing at the edge of the network, although it will require features, such as address translation, and increase the complexity. The second issue is that of effective network resource use and Ethernet’s lack of load balancing. In the LANs, where link utilization is low with a tree-like logical layer (IEEE Standard 802.1D), simple routing with possible bottlenecks still give acceptable performance. However, MANs will need to support many multiplexed sources in the core and will require more flexible routing algorithms to efficiently distribute loads over alternative paths to maximize available capacity. Such load distribution and predictable routing can be achieved through the explicit teaching (rather than passive learning) of MAC table configuration. This would also mean a more centralized precalculated routing strategy would be required. This can be accommodated using advanced optical routing architectures, proposed in Section III and compared in Section IV. Service-level guarantees: In the current LANs, overprovisioning and a relatively small number of users has meant that perceived levels of service are acceptable; however, this is not the case for metroscale networks. In core networks, careful policing of resources is required to prevent flows from interfering with each other and monopolizing resources. Soft guarantees may be achieved by careful dimensioning, routing, and traffic prioritization, but this requires complicated planning, management, and control of traffic sources, which would have to be done in the electrical layer (see RPR subsequently in the paper). Alternatively, harder guarantees can be achieved through the use of physical-layer constraints, such as the timeslot architectures proposed in Section III. Resilience: In the case of node or link failure, the network must quickly recognize this and reconfigure to be able to continue transporting data. Current Ethernet spanning tree reconfigurations are too slow (IEEE Standard 802.1D takes 1 min while 802.1w takes 2–3 s [12]) to provide the millisecond-scale protection switching expected of metro core networks. Existing metro Ethernet solutions like RPR provide protection against failure in less than 50 ms, in accordance with most telecom equipment specifications. This is sufficient for most current applications, although all dynamic routing algorithms must incorporate link failure restoration mechanisms and associated wavelength requirements. OAM and signaling: Management and monitoring of Ethernet network devices is currently done through simple network

2422

management protocol (SNMP) and remote monitoring protocol (RMON) requests that are sent in-band within a layer 3 protocol packet (e.g., IP), and as such, the management is part of the actual data flow. Ideally, the data carried will not interfere with OAM signaling. A possible solution to this is to use special new frame types (in a similar way to PAUSE frames in IEEE Standard 802.3x flow control) that can be sent between adjacent switches. Solutions to some of these problems already exist: the rapid ring spanning tree protocol (RRSTP) [13] is a proprietary spanning-tree replacement that attempts to provide protection switching (in about 400 ms) through the use of ring-based logical topologies. It provides little in the way of solutions for the other problems. A more complete attempt at providing metro-class Ethernet connectivity is RPR [4], [14], which uses a static ring of point-to-point fibers and sends Ethernet-like (frames are not directly compatible but are translatable to/from the original Ethernet format) frames around the ring. The use of a ring logical topology allows for fast protection switching (50 ms), while careful output queuing at the nodes provides fairness in resource use as well as a number of service quality classes. However, RPR uses hop-by-hop forwarding and optical links as channels only, not using optical routing or end-to-end wavelength connections, and is heavily dependent on electronic logic speeds. Another issue with the use of Ethernet in the core and at high bit rates is that of frame size limitations. The original Ethernet standard allowed frames to be a maximum of 1500 B to minimize collisions in shared-bus networks. In new applications such as storage-area networks (SANs), it would be beneficial to be able to support much larger frame sizes: IPv4 supports a maximum packet size of 64 kB. To this end, 9000-B frames have been proposed (known as jumbo frames [15]) to minimize the need for fragmentation and maximize throughput efficiency. These challenges aside, the key factor in implementing Ethernet in the core is the optical architecture and the balance between optical-layer capability and layer 2 functions, such as whether protection should be performed in the optical layer or through some Ethernet extension or function. Similarly, should entire wavelength channels be dedicated to a point-to-point Ethernet link with soft QoS guarantees performed by queuing at the Ethernet switch or should the wavelength be time-division multiplexed and provide smaller, more manageable bandwidth units with hard guarantees on QoS? The answer depends on the exact requirements of the network and the electrical or optical equipment capabilities at the required line speeds. At one extreme are optical packet networks, where the wavelengths are simple static point-to-point lightpaths (wavelength channels) with the entire capacity of the wavelength channel used for the link. RPR is an example of such a system. At the other extreme are time-slotted architectures where the optical equipment must differentiate between channels (timeslots) and forward wavelengths accordingly. The static wavelength approach, while having simpler optics, requires electronics at least as fast as the line speed ( 100 Gb/s) for MAC lookups and packet queuing. Time-slotted architectures, on the other hand, reconfigure the entire network at the end of every timeslot and need sophisticated processing (e.g., packet and slot scheduling, packet buffering, and wavelength

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Fig. 1. Static lightpath allocation in a four-node ring.

reconfiguration). The advantage, however, is in hard QoS guarantees and a finer granularity of capacity that can be dynamically grouped to adapt to demand. In the next section, we will describe the network architectures considered in this study. In Section IV, we show how the hard QoS guarantees can be combined with efficient resource usage in dynamic ring-network architectures. III. DESCRIPTION OF METRO-RING ARCHITECTURES Four sample optical metro-ring architectures were considered in this paper, covering the range from completely static and slotted ring architectures to more advanced dynamic optical-burst-switched architectures, such as the following: 1) static wavelength-routed optical network (WRON) ring; 2) static slotted ring; 3) dynamic optical burst switching with a “just-enoughtime” signaling mechanism (OBS-JET) ring; 4) dynamic wavelength-routed optical-burst-switching (WR-OBS) ring. In line with conventional ring architectures, we have assumed that in all these architectures, traffic to the destinations outside of the ring network is routed via a single node, designated as the hub, as shown in Fig. 1 for a four-node sample ring network. Architectures 1) and 2) correspond to static all-optical networks where the traffic matrix is known a priori and network resources are allocated accordingly before the start of the network operation. Architecture 1) allocates resources at the lightpath level while 2) multiplexes wavelengths between different connections, allowing a finer granularity than 1) in the resource allocation. Although this type of static network has zero blocking and is relatively simple to design and operate, its shortcomings are in either underutilizing resources at low loads (architecture 1)) or in maximum throughput achievable only for a reduced traffic load range (architecture 2)). To evaluate the potential benefits of operating all-optical networks serving as the Ethernet backbone in MAN applications dynamically, more advanced architectures 3) and 4) were also considered in this paper. Both are burst-switched networks, but

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

2423

3) uses a one-way resource reservation mechanism while 4) utilizes an end-to-end reservation scheme with acknowledgment. Architecture 1) is the type of optical system that would be commonly used with Ethernet variants like RPR where the intelligence is not in the static optical layer but rather in the Ethernet layer. Architecture 2) is very common for the metro wavelengthdivision-multiplexing (WDM) ring networks already proposed [16], and it can be seen as an example of a network with more intelligence in the optical layer. Architectures 3) and 4) extend this optical-layer intelligence to allow a traffic-demand-driven network reconfiguration. A. Static WRON Ring In a static WRON, lightpaths between network node pairs are allocated to accommodate the traffic demand matrix so wavelength collisions do not occur in the same fiber and the minimum number of wavelengths is used. Generally with this architecture, it is assumed that the traffic matrix does not change frequently and, before the network operation starts, lightpath allocation is performed and switches are configured accordingly. During network operation, data arriving at the electrical interface of an edge node is classified per destination, converted into the optical domain and allocated to a lightpath as shown in Fig. 1 (a lightpath allocation for a four-node ring for all node pairs using three wavelengths). Because the task of allocating lightpaths in a static network is an NP-problem [17], heuristics are used to perform lightpath allocation in networks of practical interest. The heuristic proposed in [18] was used here because it has been shown to achieve near-optimal results in a wide range of topologies. For the case of rings, the maximum and average wavelength reand quirement per link using this heuristic is given by , respectively (where is the number of nodes) [18], [19]. These expressions are obtained assuming that one lightpath is required between every pair of nodes and that the network is equipped with one bidirectional fiber link transmitters and receivers between adjacent nodes and per node. B. Static Slotted Ring This architecture considers unidirectional links between nodes and allocates resources using a finer granularity than a static WRON. To do so, lightpaths are not permanently established between all pairs of nodes. Instead, wavelengths are slotted and multiplexed in time over the different connections (node pairs). The wavelength/slot assignment scheme considered here is such that data directed to node must be transmitted (the wavelength serves as a receiver address), which using is slotted according to the traffic matrix. Thus, an -node ring wavelengths and, under the assumption of uniform requires slots. Fig. 2 shows the operation traffic, is divided in of this architecture under uniform traffic for a six-node ring for wavelength . Node can transmit data to node 6 only during slot . When slot finishes, node refrains from data until slot starts again. Each transmission in wavelength subslots slot could be subdivided further into

Fig. 2. Six-node ring operating as a slotted architecture, with one slot per destination (m = 1). The example shown is for transmission from any node to node 6 using wavelength  .

for higher granularity. Assuming equidistant nodes on the ring (in bits) can be calculated as network, the sub-slot size

where denotes the number of subslots, is the network diamthe velocity eter, the bit rate, the number of nodes, and of light in the optical fiber. The architecture requires tunable transmitters (e.g., fast tunable lasers [20] or laser arrays [21]) and receivers with fixed filters that can be implemented using components available today [this configuration is usually referred to as the tunable transmitter/fixed receiver (TT-FR)]. Although a laser array could be used as a transmitter, this is not necessary as slots are scheduled to avoid simultaneous transmission. In addition, for the slotted ring network the tunable laser represents the most efficient solution since 100 distributed feedback (DFB) lasers would be required for operation across the band, assuming 50-GHz channel spacing (as detailed in Section V). With reference to the static WRON, it should be noted that the slotted ring network is equipped with significantly less total ) and a significantly capacity (reduced in a factor of lower number of transmitters and receivers per node (1 versus ). This decreases the amount of input traffic that can be accommodated, as shown in Section IV. This particular slotted architecture was selected because it is representative of the type of metro-ring networks proposed in the last ten years [16]. Examples of unidirectional rings configured as TT-FR can be found in [21]–[29], with several corresponding to demonstrators such as RINGO [21], MAWSON [27], [28], and HORNET [29]. While the proposals using this architecture differ in the MAC protocol (preallocated slots, dynamically assigned slots, slot sensing, or random-access techniques), we have chosen to use a network where slots are preallocated. This ensures the ideal performance for a static traffic matrix.

2424

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Fig. 3. Slot size as multiples of Ethernet frames (9000 B) for the slotted ring 100 subslots as a function of the number of nodes and the network for network diameter. The dashed line denotes a diameter of 150 km as considered throughout this paper.

m=

For compatibility with an Ethernet environment, it is proposed that the slot size is compatible with Ethernet frame sizes (i.e., 9000-B). Fig. 3 shows the slot size as a function of the network diameter (0–200 km) and the number of nodes (2–20) for 100 subslots. The value for a bit rate of 100 Gb/s and was chosen such that for the specified network diameter of 150 km, the minimum slot size would be larger or equal 20 nodes. The overhead for all to one Ethernet frame, for burst and slot sizes in this paper was assumed to be 70 ns for transmitter tuning and receiver clock recovery (details in Section V).

Fig. 4.

Six-node OBS-JET network architecture.

Fig. 5.

Six-node WR-OBS network architecture.

C. Dynamic OBS-JET Ring OBS-JET [30] consists of sending bursts of information (electronically built from packets arriving at the edge of the network) through the optical core after a control packet has configured switches in a hop-by-hop basis (see Fig. 4). Bursts must remain in the optical domain once released, and the core nodes do not have optical buffering capability. For this reason, the burst is kept for a period of time (called offset time) in the electronic buffer of the edge node, chosen to be long enough to allow for switches in the path to be configured by the control packet when the burst arrives. As bursts are assumed to be in the range of tens of kilobytes, there is no time for end-to-end path reservation. As a result, bursts can be dropped at any point along the path to the destination due to contention. To decrease the probability of contention, full wavelength conversion is compulsory in every node of the network; hence, wavelengths are used only as high-speed transmission channels rather than for routing. D. Dynamic WR-OBS Ring In WR-OBS [31], as in OBS-JET, packets are electronically aggregated into bursts at the edge of the network according to their destination. Unlike conventional OBS, however, end-to-end lightpath reservation is required before sending a burst through the optical core (see Fig. 5). To do so, at some point of the aggregation process, a request is sent to the network

control and management node to find and reserve resources for the burst. Once the lightpath has been reserved in the core, an acknowledgment with information on the reserved lightpath is sent to the edge node and then the burst can be transmitted. If a lightpath is not found, a negative acknowledgment is sent to the edge node, and the burst is dropped. End-to-end lightpath reservation means that bursts must be in the millisecond range

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

(to allow time for lightpath acknowledgment, mainly determined by propagation delays) and that hard guarantees for QoS requirements such as latency can be provided. WR-OBS can be designed in several ways according to the burst assembly mechanism and the routing and wavelength allocation (RWA) algorithm used. In this paper, a centralized WR-OBS architecture [32] is considered where the lightpath scheduler resides in a control node equipped with a “re-attempt” function. This means that if a request cannot successfully reserve a lightpath, it remains in the control node until resources become available or until a deadline expires [33], when the request is rejected. This deadline, called maximum scheduling time is a key in providing end-to-end delay guarantees and decreasing the blocking probability. IV. OPTICAL METRO RINGS: PERFORMANCE COMPARISON The four ring architectures were evaluated and compared in terms of throughput, mean end-to-end delay, and required capacity in the context of the metro Ethernet requirements listed in Section II. A. Modeling Assumptions Ten assumptions were made in this paper. 1) Nodes, ranging from 8 to 20, are equally spaced around rings of 150-km diameter. 2) Uniform traffic distribution is used (the results of nonuniform traffic distribution can be found in Section IV-C). 3) Poisson-distributed arrivals of fixed-size data units (packets/frames in static networks and bursts in dynamic ones) are used, allowing for analytically tractable treatment. For static networks this is, however, an optimistic assumption as the size of arriving data units would generally not be fixed. For burst-switched architectures instead, the interarrival time of fixed-size bursts generated from packets arriving as a Poisson process is a Gamma distribution with a coefficient of variation 1.0. The assumption of a Poisson-distributed arrival process for burst-switched networks is, therefore, a worst-case scenario. For the slotted ring, it is assumed that a data unit fits exactly within one slot. 4) The burst transmission time is at least as long as the time required to allocate resources (efficiency criterion). This results in shorter bursts for OBS-JET (between 2 and 25 kB, approximately) than for WR-OBS (25 MB, approximately). Notice that, especially at low loads, there is an inevitable tradeoff between maximizing the throughput (hereafter referred to as efficiency criterion) and guaranteeing end-to-end delay (hereafter referred to as latency criterion). The efficiency criterion demands that the aggregation process continues until the minimum burst size is achieved to ensure minimum resource utilization. At low loads, however, this may lead to unacceptable delays. In such a situation, the latency criterion demands that a burst is sent irrespective of its current size to guarantee that the packets contained in the burst meet prespecified end-to-end delay deadlines (as

2425

applied, for instance, in [32]). This may result in reduced throughput. The tradeoff between the latency and the efficiency criterion will be investigated in more detail later in this section. 5) Buffers are sufficiently large to neglect overflow effects. 6) The maximum input bit rate to buffers and maximum bit rate per wavelength are both 100 Gb/s. 7) In the OBS-JET ring, the shortest path (SP) is used to choose the paths and the first-fit with void filling (FF-VF) [34] is used to allocate wavelengths. 8) In the WR-OBS ring, the shortest path/first fit (SP-FF) is used to allocate lightpaths to requests; although this results in a higher blocking probability than for more complex algorithms, SP-FF was used because it is fast and allows a comparison with the RWA algorithm used in OBS-JET networks. 9) The lightpath request in WR-OBS is sent as soon as the first data unit arrives to the edge buffer of 36 ms was considered for 10) A maximum WR-OBS rings. In a WR-OBS ring configured to meet an end-to-end delay of 40 ms (significantly less than the limit of 100 ms established for time-critical applications [35]), the difference of 4 ms would account for propagation delays. B. Performance Metrics and Results The network performance in the different architectures was compared using the following performance metrics and extending work in [36]. 1) Network Throughput: The throughput is defined here as the mean value of the traffic load successfully delivered from source to destination, for all nodes. For the static WRON, a maximum bandwidth of bits per second (wavelength bit rate) is allocated per connection, while bits the slotted ring network is dimensioned to carry to per second. As long as the input data rate does not exceed these bandwidth limitations, these networks are loss free. Therefore

where is the offered load per connection. Notice the limited load range of the slotted ring compared with the static ring. In contrast, in dynamic networks, resources are not preallocated to connections and, hence, contention may arise during network operation. This reduces the throughput due to bursts dropped (OBS-JET) or blocked (WR-OBS) to

where is the offered load and is the probability of data not delivered at the destination node. was evaluated through simulation. To validate the simulator, we compared simulation results with ones calculated analytically using the Erlang fixed-point model [37] for the OBS-JET ring simulations and the model presented in [38] for ). The comparison was WR-OBS rings (with performed for small-size rings (where the assumptions of the

2426

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

However, as the number of nodes increases, the performance difference between the investigated architectures reduces because of higher capacity provisioning. 2) Delay: The next performance parameter investigated was the mean end-to-end delay, denoted by , which corresponds to the mean time elapsed since the arrival of the first bit of a data unit at the transmission buffer until its successful reception at the destination node, that is

where is the time that a data unit spends in the source is the bit transtransmission buffer before being transmitted, mission time, and is the propagation time from the source ; hence, this to the destination node. At 100 Gb/s, time was neglected in the evaluation of . is calculated modeling For the case of a static WRON, each node buffer as an M/D/1 continuous queue [39] with ser(where is the data unit size vice time equal to in bits and the wavelength bit rate). The propagation time is , where is the mean path length in given by the WRON ring and is the speed of light in fiber. is given by

Fig. 6. (a) Throughput for eight-node ring (maximum nine wavelengths per link) and (b) throughput for 20-node ring (maximum 53 wavelengths per link).

analytical models hold) and gave good agreement between simulation and analytical results. Fig. 6(a) and (b) shows the calculated throughput for each architecture, as a function of the offered load, for rings of eight and 20 nodes, respectively. For a fair comparison, dynamic rings were equipped with the same capacity required in the static WRON ring. The results show that the static WRON rings exhibit the maximum achievable throughput followed very closely by WR-OBS. In spite of full wavelength conversion, OBS-JET exhibits a performance worse than WR-OBS, which highlights over wavelength converthe benefits of introducing sion, especially at high loads. The slotted ring also achieves maximum throughput but only over a very limited load range (traffic load 0.14 and 0.05 for eight- and 20-node rings, respectively), due to the reduced bandwidth per connection. Results (not shown here) were also obtained for rings of 12 and 16 nodes (maximum 19 and 33 wavelengths per link, respectively), and the relative performance remained the same.

if

is even

if

is odd

where is the number of nodes, and the diameter of the ring. was calculated by modIn the slotted ring architecture eling each node buffer as an M/D/1 discrete queue [40] with . The propagation time is service time equal to , where begiven by cause of the use of unidirectional links in slotted ring (instead of bidirectional ones in static WRON). corresponds to , where In OBS-JET, is the mean burst aggregation time (dependent on the burst aggregation mechanism, for instance, used in [41]–[44]) and the offset time. For the aggregation mechanism considered here , where is the size (fixed-size bursts), of bursts, the bit rate, and the offered load. The offset time is given by the product of the time required to process the control packet per node and the number of nodes within the path. The time to process a control packet [to perform the wavelength ] is dominated by the memory access time (we lookup consider the use of an SRAM, with memory access times in the order of 1 ns as in current Pentium 4 processors [45]) and switch configuration time (in the order of 1 s, using three-dimensional fast microelectromechanical systems (MEMS) [46]). The prop, where agation time is given by . is given by Finally, in WR-OBS , where is the mean burst is the time elapsed since the arrival of aggregation time, the first packet of a burst until the lightpath request is sent to the is the control node (assumed here as equal to zero), mean time for the request to be propagated to the control node is the mean time the request and back to the edge node, and remains in the control node to be allocated a lightpath (with an

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

upper bound equal to ). The propagation time is given , . as Fig. 7(a) and (b) shows the calculated end-to-end delay as a function of the offered load for the eight- and 20-node rings, respectively. As before, the dynamic rings were equipped with the same capacity as the static ring. For static and slotted rings, the data unit size considered was equal to one frame. According to the efficiency criterion, bursts of about 2 and 25 kB were used for the OBS-JET rings of eight and 20 nodes, respectively. For WR-OBS rings, bursts of about 25 MB were used. The mean was time requests spent in the control node in WR-OBS obtained through simulation. It can be seen that the static WRON architecture exhibits the lowest end-to-end delay. This is dominated by the propagation time and therefore does not change significantly with the number of nodes (as the network diameter was kept constant). The slotted ring can operate only at very low loads, and hence only one point is shown at the maximum load of 0.14 and 0.05 for the eightand 20-node rings, respectively. Slotted ring delay is mainly due to the propagation times, and therefore it is higher than WRON delay due to the use of unidirectional links. OBS-JET delay values are very similar to those of static WRON for the eight-node ring. For rings with higher node count, the OBS-JET delay increases slightly as paths with more hops result in higher offset time. This effect results in longer aggregation times, applying the efficiency criterion as defined in assumption 4); hence, bursts become longer for the 20-node ring. However, due to the low switching times considered ( 1 s), the delay introduced by the offset time is still negligible. WR-OBS exhibits the longest mean end-to-end delay due to the end-to-end reservation process, the aggregation mechanism (fixed-size bursts due to the efficiency criterion), and the processing time in the central node. At low traffic loads, it may take considerable time (tens of milliseconds) for a sufficiently large burst to build up in the edge router buffer, in order to allow high wavelength utilization, while at high loads the processing time per lightpath request in the central node (with upper-bound ) contributes the most. For the considered load equal to range (0.1–0.9), the end-to-end delay for all rings remains under 40 ms, well under the limit of 100 ms required for time-critical applications [35]. Results for rings of 12 and 16 nodes (not shown here) were also obtained, and the relative performance remained the same as shown in Fig. 7. The high delays shown in Fig. 7 can be avoided, however, if bursts are released into the core node as soon as a deadline expires (latency criterion), instead of waiting for a minimum burst size to build up (efficiency criterion). This deadline-based operation may lead to reduced throughput, however, highlighting a tradeoff between latency and throughput. To understand this, we now study the tradeoff between delay and load for an eight-node ring, which was configured to guarantee a maximum end-to-end delay of 40 ms. To do so, the maximum scheduling time at the control node was again set to 36 ms for a maximum 4-ms propagation delay and processing in the control node, but the burst is released into the optical core as soon as the acknowledgment from the control node is received at the edge node. Hence, bursts of smaller size than required by the efficiency criterion may be sent, but a maximum end-to-end delay is guaranteed. Fig. 8 shows the results for throughput and end-to-end delay for an

2427

Fig. 7. (a) Mean end-to-end delay versus offered load for eight-node ring and (b) mean end-to-end delay versus offered load for 20-node ring.

eight-node ring configured as WR-OBS for the cases when the efficiency (lines with circles) or the latency criterion (lines with squares) is applied. It can be seen that high resource utilization with near-optimal throughput performance can be achieved in the case of the efficiency criterion, but at the expense of delays as high as several tens of milliseconds when operating at low loads ( 0.3). With the deadline-based operation, the delay is much lower than the end-to-end delay target of 40 ms, but now the throughput performance degrades at high loads ( 0.7). From Fig. 8, it can be seen that there is a value of traffic load at which the network must switch between the two burst aggregation schemes (efficiency and latency criterion) in order to achieve the best compromise between throughput and delay. In the case of the sample traffic used in this paper, this transfer happens at a value for the traffic load of 0.7. For traffic loads under this value, the best compromise between throughput and latency is given by the latency criterion, while the efficiency criterion is the best choice for traffic loads in excess of 0.7.

2428

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Fig. 8. Comparison of end-to-end delay and throughput of WR-OBS ring when applying the latency and the efficiency criterion for burst aggregation.

3) Capacity: The capacity required to satisfy the traffic demand in each of the architectures considered has been normalized to the capacity required in the WRON architecture, on a link-by-link basis, for all links. The capacity required for OBS-JET and WR-OBS was calculated under the requirement , sufficient of a maximum blocking probability lower than to ensure acceptable loss rates according to [35]. Fig. 9(a) and (b) shows the calculated capacity for networks with eight and 20 nodes, respectively. From the results, it can be seen that the WR-OBS network architecture requires significantly lower capacity than OBS-JET to achieve the same blocking probability and fewer wavelengths than the static WRON (and slotted ring) for loads up to 0.9. These results, along with those of delay and throughput, show that the WR-OBS architecture combines high throughput (comparable with static end-to-end connections) with low delay ( 40 ms for loads higher than 0.1) while requiring a minimum number of wavelengths, making it potentially the most cost-efficient architecture to deploy. Only the static WRON exhibits better performance in terms of throughput and delay, but it requires more wavelengths at low loads. C. Impact of Nonuniform Traffic While much research on network performance has been carried out under the assumption of uniform traffic matrices, realworld traffic distribution is highly nonuniform [47], [48]. This has implications on the amount of network resources required to support a given demand. In this section, the impact of nonuniform traffic matrices applied to static WRON and WR-OBS ring network architectures is investigated in terms of the required capacity. We are no longer considering the slotted ring and OBS-JET architectures since, under uniform traffic conditions, their performance is suboptimal, as shown in the previous section. This choice is also supported by results reported in [49], where end-to-end versus hop-by-hop reservation protocols were analyzed (in the context of ATM networks). The relative performance did not change when switching from uniform to nonuniform traffic.

Fig. 9. (a) Required capacity versus offered load for eight-node ring and (b) required capacity versus offered load for 20-node ring, normalized to the capacity required in the WRON architecture.

We apply the following model to describe nonuniformity, while, at the same time, allowing a comparison to the performance of a network operating under uniform input traffic conditions. The total traffic load of the nonuniform traffic matrix is assumed to be the same as in the uniform case, equal to , where is the number of nodes and is the . The type of (uniform) offered load per node pair nonuniformity assumed is such that there is one node in the network, the hub , which is used as a sink for a fraction of the total network load. The remaining is uniformly distributed among the rest of the nodes. Therefore, traffic (for uniform traffic) and (for nonuniform matrices traffic) are defined as shown in Fig. 10, where and . The level of nonuniformity varies from (uniform traffic case, ) to 1, where the hub concentrates the entire network traffic load.

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

Fig. 10.

Fig. 11.

Definition of uniform and nonuniform traffic matrices.

Required capacity versus level of nonuniformity for eight-node ring.

By way of an example, Fig. 11 shows the capacity requirement as a function of the level of nonuniformity for a total for an eight-node ring. In the graph, traffic load of ranges from 0.25 (uniform case) to 1 (all traffic load concentrated in the hub node). The required capacity has been normalized to the capacity required in the WRON architecture under uniform traffic, and the capacity required for WR-OBS was calculated under the requirement of a maximum blocking proba. bility lower than It can be seen that the capacity required in the static ring increases with the level of nonuniformity , except in the exwhen all traffic concentrates in some links, treme case of leaving others unused. The dynamic WR-OBS ring, however, required a capacity very close to 1 for all levels of uniformity, highlighting the potential of this architecture as a cost-effective and robust alternative. V. IMPLEMENTATION OF 100 GbME The extension of the existing Ethernet standards toward 100 GbME also raises a number of questions with respect to the physical implementation of such a scheme, and these are addressed here. Ethernet has traditionally depended on the availability of large amounts of unallocated bandwidth, resulting in a network infrastructure that was easy to manage but capacity inefficient.

2429

With channel data rates approaching 80 160 Gb/s, as envisaged for an implementation of 100 GbME, this assumption will no longer hold. Although the bandwidth of optical fiber exceeds tens of terahertz (from 1200–1700 nm), the usable part of this spectrum may be, for some time to come, limited by the bandwidth of erbium-doped optical fiber amplifiers (EDFA), typically covering the range 1530–1565 nm, roughly equivalent to a bandwidth of 4 THz. Considering standard modulation formats for optical networks like the nonreturn-to-zero (NRZ) and return-to-zero (RZ), the channel spacing required for the envisaged data rates of 80 160 Gb/s would be 200 GHz minimum. Combined with the available EDFA bandwidth of 4 THz and wavelength requirements for static end-to-end connections of 50 wavelengths for a 20-node network (based on capacity calculations in Section IV-B), we note that the EDFA bandwidth is likely to prove insufficient to support this architecture for the following reasons. 1) Bandwidth in 100 GbME may become scarce when using standard telecom transmission fiber and technology (e.g., EDFAs), requiring additional fibers and amplifiers, or more complex amplifier technology such as -band amplifiers [50]. A more detailed discussion of the spectral efficiency of NRZ and RZ modulation and formats in the context of 100 GbME networks can be found in [51]. 2) Wavelength savings through dynamic allocation becomes key for the operation of future MANs, motivating the need for a capacity management component in the MAC protocol as discussed in Section II and dynamic network architectures as described in Section III. 3) 100 GbME would, therefore, benefit from the application of advanced modulation formats such as carrier-suppressed RZ (CS-RZ) as a spectrally efficient format for amplitude modulation or differential quadrature phaseshift keying (dQPSK) for phase modulation techniques [52]. Although it might be difficult to implement these advanced modulation techniques directly at high bit rates of 80 160 Gb/s, they could be useful when used as subchannels in a waveband format, as discussed hereafter. A key challenge in the design of 100 GbME is thus in the definition of the PHY for operation at high bit rates. Although integrated circuits (ICs) have already been demonstrated to be—in principle—operational for selected functionalities of electrical signals up to 100 Gb/s [53], [54], obstacles such as lack of amplification or modulation and bonding problems still prevail. It is, therefore, believed that in 100 GbME, all signals will be optical as direct modulation at 100 Gb/s is yet impractical, and a two-stage multiplexing technique is proposed, as shown in Fig. 12(a) for WDM and Fig. 12(b) for optical time-division multiplexing (OTDM). It is assumed in all cases that the minimum input bit rate into the multiplexer would be in the form of GbE signals. At the first multiplexing stage, ten of these signals would be electrically aggregated into a single 10-GbE stream, for which the technology already exists today. The novelty of the approach discussed here is to add a second optical multiplexing stage, which aggregates ten 10-GbE signals into the 100-GbME signal, via an optical multiplexer. This multiplexer

2430

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Fig. 12. Principle of the envisaged two-stage multiplexing from lower bit rate signals (from GbE to 10 GbE and from 10 GbE to 100 GbE) for (a) WDM and (b) OTDM.

Fig. 13.

Network node based on a waveband approach (densely spaced WDM).

could be based on WDM (which could be based on existing PHY of IEEE Standard 802.3ae) or OTDM. The proposed WDM node is shown in Fig. 13 and consists of the multiplexing stage as described previously, as well as an optical multiplexing/demultiplexing unit and an switch for connection to the ring network. The optical multiplexer and demultiplexer are used to add/drop management in. If formation on the control channel with wavelength assofixed wavelength addressing is used, the waveband ciated with the node would be dropped, while all other wave-

switch is used to add lengths bypass the node. A . All wavebands are aginformation to any other waveband gregated together with the control channel at the output of the node. The OTDM solution is shown in Fig. 14; the operational principle is the same as for the WDM source. The main difference in the transmitter and receiver are the optical multiplexing and demultiplexing units. The optical multiplexer would con5 ps and separate delay sist of a short pulse source lines for multiplexing lower bit rate signals into the 100-GbME

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

Fig. 14.

2431

Network node based on OTDM technique.

stream. The key challenge of this scheme is the provisioning of 1 s , which generates suffia rapidly tunable pulse source ciently short pulses [55]. The components and devices used for future dynamic network application will introduce an inevitable overhead for the operation of the network. The main contributions arise from the laser tuning times at the transmitter and clock recovery at the receiver. For fast tunable lasers, the switching between two lasers can be achieved in typically 50 ns when no active control is used, although it can be reduced to 5–10 ns when using active wavelength stabilization [56]. Clock recovery is a critical function within the edge router receiver. Passive radio-frequency (RF) microwave bandpass filters with high factors are used for clock recovery since the acquisition time of phase-locked loops (PLLs) with approximately 500 ns is typically too large for fast burst and packet-switching operations [57]. When using passband filters, however, there is an inevitable tradeoff between the bandwidth of the filter, the rise time, and jitter (see, e.g., [58]). As an example, at 10 Gb/s, a filter with a 50-MHz passband (3 dB) would show a rise time of approximately 20 ns. As a worst-case estimate, we assumed that the total overhead incurred would be approximately 70 ns.

VI. SUMMARY AND CONCLUSION This paper discussed key aspects for the design and deployment of 100 GbME and the necessary modifications to enable its extension from LANs to MAN rings. This requires Ethernet to include network management functionality currently not required in the LAN and to adapt the frame size to make the best use of the high bit rates, while providing backward compatibility. Four different optical ring network architectures that could serve as a potential future Ethernet backbone under the 100 GbME were analyzed and their performance compared in terms of defined performance metrics, namely throughput,

delay, and number of required wavelengths. It was shown that the highest throughput is achieved in the static WRON followed very closely by WR-OBS, but the WR-OBS architecture provided the largest wavelength savings for network loads less than 0.9. However, the WR-OBS trades off throughput for latency, reaching end-to-end delays of between 6 and 40 ms (as opposed to 0.7 ms for the static WRON ). The delay could be reduced by using a deadline-driven aggregation mechanism at the edge node but at the expense of reduced throughput at very high loads. However, it was also shown that both low delay and high throughput can be achieved by selecting the appropriate aggregation mechanism (deadline or burst-size driven) as a function of the input traffic load. Under nonuniform traffic, the WR-OBS architecture also offered wavelength savings with respect to static WRON. Since the WR-OBS architecture combines high throughput (comparable to static end-to-end connections) with acceptable delay ( 40 ms for loads higher than 0.1 for a burst-size-driven aggregation mechanism, which is much lower than the limit of 100 ms imposed to time-critical applications) while requiring a minimum number of wavelengths, it would potentially be the most cost-efficient architecture to deploy. The high capacity per channel in a ring configuration requires a bandwidth-limitation-aware MAC protocol when operating over a single-fiber/EDFA amplification infrastructure to accommodate all channels. Further wavelength savings could be achieved by the use of advanced modulation formats with high spectral efficiency ( 1 b/s/Hz) to allow cost-efficient operation within the band only, and possible physical implementation network nodes (edge and core) have been proposed. The results of this study explain the requirements for the nextgeneration metropolitan network architectures and show the potential of extending the Ethernet to operate in a metropolitan environment. This paper shows the important role that dynamic OBS network architectures are likely to play in implementation of future networks and also points to key issues that must be addressed in the future standardization process.

2432

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

ACKNOWLEDGMENT A. Zapata would like to thank Universidad Tecnica Federico Santa Maria for support during her leave of absence. REFERENCES [1] S. Keshav, An Engineering Approach to Computer Networking. Reading, MA: Addison-Wesley, 1997. [2] N. Ghani, J.-Y. Pan, and X. Cheng, “Metropolitan area networks,” in Optical Fiber Communications—Vol. IV-B: Systems and Impairments, I. P. Kaminow and T. Li, Eds. San Diego, CA: Academic, 2002, pp. 329–403. [3] 10 Gb/s Ethernet, IEEE Standard 802.3ae, 2002. [4] Resilient Packet Ring (RPR) Alliance, IEEE Standard 802.3ae, 2003. [5] MAC Bridges, IEEE Standard 802.1D, 1993. [6] Link Aggregation, IEEE Standard 802.3ad, 2000. [7] GARP/GMRP Timer Values, IEEE Standard 802.1p, 1997. [8] Flow Control in Full-Duplex LANs, IEEEE Standard 802.3x, 1999. [9] VLAN TAG, IEEE Standard 802.3ac, 1998. [10] Multiple Spanning Trees, IEEE Standard 802.1s, 2002. [11] J. Spencer, M. Düser, A. Zapata, I. de Miguel, P. Bayvel, D. Breuer, N. Hanik, and A. Gladisch, “Design considerations for 100-Gigabit metro Ethernet (100 GbME),” in Proc. 29th Eur. Conf. Optical Commun. (ECOC 2003), vol. 4, Rimini, Italy, Sept. 21–25, 2003, pp. 904–905. [12] Rapid Reconfiguration of Spanning Tree, IEEE Standard 802.1w, 1998. [13] (2002) Rapid Ring Spanning Tree Protocol. Riverstone Networks. [Online]. Available: http://www.riverstonenet.com [14] Resilient Packet Ring, IEEE 802.17 Working Group, 2003. [15] P. Dykstra. (1999) Gigabit Ethernet Frames and Why You Should Care. [Online]. Available: http://sd.wareonearth.com/~phil/.html [16] M. Herzog, M. Maier, and M. Reisslein, “Metropolitan area packetswitched WDM networks: A survey on ring systems,” IEEE Communications Surveys, vol. 6, no. 2, pp. 2–20, 2nd Quarter 2004. [17] M. R. Garey and D. S. Johnson, Computers and Intractability: a Guide to the Theory of NP-Completeness. New York: W. H. Freeman, 1979. [18] S. Baroni and P. Bayvel, “Wavelength requirements in arbitrarily connected wavelength-routed optical networks,” J. Lightwave Technol., vol. 15, pp. 242–251, Feb. 1997. [19] S. K. Korotky, “Network global expectation model: A statistical formalism for quickly quantifying network needs and costs,” J. Lightwave Technol., vol. 22, pp. 703–722, Mar. 2004. [20] J. E. Simsarian, A. Bhardwaj, J. Gripp, K. Sherman, Y. Su, and C. Webb et al., “Fast switching characteristics of a widely tunable laser transmitter,” IEEE Photon. Technol. Lett., vol. 15, pp. 1038–1040, Aug. 2003. [21] R. Gaudino, A. Carena, V. Ferrero, A. Pozzi, V. De Feo, P. Gigante, F. Neri, and P. Poggliolini, “RINGO: A WDM ring optical packet network demonstrator,” in Proc. 27th Eur. Conf. Optical Commun. (ECOC 2001), vol. 3, Amsterdam, The Netherlands, Sept. 30–Oct. 4 2001, pp. 620–621. [22] M. A. Marsan, A. Bianco, E. Leonardo, M. Meo, and F. Neri, “MAC protocols and fairness control in WDM multirings with tunable transmitters and fixed receivers,” J. Lightwave Technol., vol. 14, pp. 1230–1244, June 1996. [23] A. Bianco, V. Distefano, A. Fumagalli, E. Leonardo, and F. Neri, “A-posteriori access strategies in all-optical slotted WDM rings,” in Proc. IEEE GLOBECOM 1998, vol. 1, Nov. 8–12, 1998, pp. 300–306. [24] K. Bengi and H. R. van As, “Efficient QoS support in a slotted multihop WDM metro ring,” IEEE J. Select. Areas Commun., vol. 20, pp. 216–227, Jan. 2002. [25] J. Fransson, M. Johansson, and M. Roughan, “Design of a medium access control protocol for a WDMA/TDMA photonic ring network,” in Proc. IEEE GLOBECOM 1998, vol. 1, Sydney, Australia, Nov. 8–12, 1998, pp. 307–312. [26] J.-Y. Wang, W.-S. Hwang, W.-F. Wnag, and C.-K. Shieh, “Modeling multi-channel slotted ring networks with tunable transmitters and fixed receivers,” in Proc. IEEE Int. Conf. Networks (ICON 2000), Singapore, Singapore, Sept. 5–8, 2000, pp. 449–456. [27] M. A. Summerfield, “Wavelength division multiplexing for local area networks,” in Proc. Conf. Lasers Electro-Optics (CLEO)/Pacific Rim’99, Seoul, Korea, Aug. 30–Sept. 3, Paper FN4, pp. 1215–1216. [28] M. J. Spencer and M. A. Summerfield, “WRAP: A medium access control protocol for wavelength-routed passive optical networks,” J. Lightwave Technol., vol. 18, pp. 1657–1676, Dec. 2000. [29] D. Wonglumsom, I. M. White, K. Shrikhande, M. S. Rogge, S. M. Gemelos, A. F.-T. An, Y. Fukashiro, M. Avenarius, and L. G. Kazovsky, “Experimental demonstration of an access point for HORNET—A packet-over-WDM multiple-access MAN,” J. Lightwave Technol., vol. 18, pp. 1709–1717, Dec. 2000.

[30] C. Qiao and M. Yoo, “Optical Burst Switching (OBS)—A new paradigm for an optical Internet,” J. High Speed Networks, vol. 8, pp. 69–84, Aug. 1999. [31] M. Düser and P. Bayvel, “Analysis of wavelength-routed optical burst switched (WR-OBS) network architecture,” J. Lightwave Technol., vol. 20, pp. 574–585, Apr. 2002. [32] I. de Miguel, M. Kozlovski, and P. Bayvel, “Provision of end-to-end delay guarantees in wavelength-routed optical burst switched networks,” in Next Generation Optical Network Design and Modeling, A. Bianco and F. Neri, Eds. Norwell, MA: Kluwer, 2003, pp. 85–100. [33] E. Kozlovski, M. Dueser, I. de Miguel, and P. Bayvel, “Analysis of burst scheduling for dynamic wavelength assignment in optical burst switched networks,” presented at the 14th Annu. Meeting IEEE Lasers & ElectroOptics Society (LEOS 2001), San Diego, CA, Nov. 12–13, 2001, Paper TuD2. [34] Y. Xiong, M. Vandenhoute, and C. H. Cankaya, “Control architecture in optical burst-switched WDM networks,” IEEE J. Select. Areas Commun., vol. 18, pp. 1838–1851, Oct. 2000. [35] N. Seitz, “ITU-T QoS standards for IP-based networks,” IEEE Commun. Mag., vol. 41, pp. 82–89, June 2003. [36] A. Zapata, I. de Miguel, M. Düser, J. Spencer, P. Bayvel, D. Breuer, N. Hanik, and A. Gladisch, “Performance comparison of static and dynamic optical metro ring network architectures,” in Proc. 29th Eur. Conf. Optical Communication (ECOC 2003), vol. 3, Rimini, Italy, Sept. 21–25, 2003, pp. 374–375. [37] F. P. Kelly, “Loss networks,” Annals Appl. Prob., vol. 1, no. 3, pp. 319–378, Aug. 1991. [38] A. Mokhtar and M. Azizo˘glu, “Adaptive wavelength routing in all-optical networks,” IEEE/ACM Trans. Networking, vol. 6, pp. 197–206, Apr. 1998. [39] L. Kleinrock, Queueing Theory. New York: Wiley, 1975, vol. I. [40] O. Yang, “Performance comparison of some discrete service time systems,” Perform. Eval., vol. 23, pp. 261–284, 1995. [41] A. Ge, F. Callegati, and L. Tamil, “On optical burst switching and selfsimilar traffic,” IEEE Commun. Lett., vol. 4, pp. 98–100, Mar. 2000. [42] G. Hu, K. Dolzer, and C. Gauger, “Does burst assembly really reduce the self-similarity?,” in Proc. Optical Fiber Communication Conf. (OFC 2003), Atlanta, GA, Mar. 2003, pp. 124–126. [43] X. Yu, Y. Chen, and C. Qiao, “A study of traffic statistics of assembled burst traffic in optical burst switched networks,” in Proc. Opticomm 2002, Boston, MA, 29 July–1 Aug. 2002, pp. 149–159. [44] I. de Miguel, M. Düser, and P. Bayvel, “Traffic load bounds for optical burst switched networks with dynamic wavelength allocation,” in Towards an Optical Internet: New Visions in Optical Network Design and Modeling, A. Jukan, Ed. Norwell, MA: Kluwer, 2002, pp. 209–224. [45] G. Hinton, D. Sager, M. Upton, D. Boggs, D. Carmean, A. Kyker, and P. Roussel. (2001, 1st Quarter) “The microarchitecture of the Pentium 4 processor”. Intel Technology J. [Online]. Available: http://www.intel.com/technology/itj/q12001/articles/art_2.htm [46] G. Bennett, “Optical switches,” presented at the Opticon 2001, San Jose, CA, Aug. 13–16, 2001. [47] W. Fang and L. Peterson, “Inter-AS traffic patterns and their implications,” in Proc. IEEE GLOBECOM 1999, vol. 3, Rio de Janeiro, Brazil, Dec. 5–9, 1999, pp. 1859–1868. [48] S. Bhattacharyya, C. Diot, J. Jetcheva, and N. Taft, “POP-level and access-link level traffic dynamics in a tier-1 POP,” presented at the ACM SIGCOMM Internet Measurement Workshop, San Francisco, CA, Nov. 2001. [49] I. Widjaja, “Performance analysis of burst admission control protocols,” in Proc. Inst. Elect. Eng.—Communications, vol. 142, Feb. 1995, pp. 7–14. [50] E. Desurvire, D. Bayart, B. Desthhieux, and S. Bigo, Erbium-Doped Fiber Amplifiers. New York: Wiley, 2002. [51] A. Zapata, M. Düser, J. Spencer, P. Bayvel, I. de Miguel, D. Breuer, N. Hanik, and A. Gladisch, “Investigation of future optical metro ring networks based on 100-gigabit metro ethernet (100 GbME),” in Proc. 4. ITG—Fachtagung Photonische Netze, Leipzig, Germany, May 5–6, 2003, pp. 89–98. [52] P. J. Winzer and R.-J. Essiambre, “Advanced optical modulation formats,” in Proc. 29th Eur. Conf. Optical Commun. (ECOC 2003), vol. 4, Rimini, Italy, Sept. 21–25, 2003, pp. 1002–1003. [53] E. Sano, “High-speed lightwave communication IC’s based on III-V compound semiconductors,” IEEE Commun. Mag., vol. 39, pp. 154–158, Jan. 2001. [54] K. Murata, K. Sano, S. Sugitani, H. Sugahara, and T. Enoki, “100 Gbit/s multiplexing and demultiplexing IC operations in InP HEMT technology,” Electron. Lett., vol. 38, pp. 1529–1531, Nov. 2002.

ZAPATA et al.: NEXT-GENERATION 100 GbME USING MULTIWAVELENGTH OPTICAL RINGS

[55] L. Davis, M. G. Young, D. Dougherty, S. Keo, R. Muller, and P. Maker, “Multiwavelength modelocked laser arrays for WDM applications,” IEE Electron. Lett., vol. 34, pp. 1858–1860, Sept. 1998. [56] C. C. Renaud, M. Dueser, C. F. C. Silva, B. Puttnam, T. Lovell, P. Bayvel, and A. J. Seeds, “Nanosecond channel-switching exact optical frequency synthesizer using an optical injection phase-locked loop (OIPLL),” IEEE Photon. Technol. Lett., vol. 16, pp. 903–905, Mar. 2004. [57] A. Ravi, G. Banerjee, R. E. Bishop, B. A. Bloechel, L. R. Carley, and K. Soumyanath, “10 GHz, 20 mW, fast locking, adaptive gain PLL’s with on-chip frequency calibration for agile frequency synthesis in a .18 m CMOS process,” in Proc. IEEE Symposium VLSI Circuits, Kyoto, Japan, June 12–14, 2003, pp. 181–183. [58] P. Monteiro, J. N. Matos, A. Gameiro, and J. R. F. da Rocha, “10 Gbit/s timing recovery circuit using dielectric resonator and active bandpass filters,” Electron. Lett., vol. 28, no. 9, pp. 819–820, Apr. 1992.

Alejandra Zapata (S’02) was born in Valparaiso, Chile, in 1972. She received the B.Eng. and M.Sc. degrees in electronic engineering from the Universidad Tecnica Federico Santa Maria (UTFSM), Valparaiso, Chile, in 1993 and 2001, respectively. Her M.Sc. thesis dealt with performance evaluation of scheduling algorithms in asynchronous transfer mode networks using Markov chains with rewards. She is currently working toward the Ph.D. degree in the Optical Network Group at the University College London, London, U.K., where her work focuses on the dynamic allocation of resources in all-optical networks, in particular the impact of dynamic routing on the performance and scalability aspects of wavelength-routed optical-burst-switched (WR-OBS) network architectures. She joined the Department of Electronics in the Universidad Tecnica Federico Santa Maria as a Junior Lecturer in 1997, where she lectured on the topics of digital systems, computer architectures, and computer networks, and worked on two research projects funded by Fondecyt, Chile, which were concerned with routing and reliability aspects of multicast communications. Ms. Zapata is a Junior Member of the British Federation of Women Graduates, United Kingdom. She received the Award to the Best Student from the UTFSM Alumni Association in 1996 granted, the Best Female Graduate of UTFSM Award from Zonta International in 1996, the M.H. Joseph Prize granted for academic excellence in the field of engineering from the British Federation of Women Graduates, United Kingdom, in 2003, and an IEEE Lasers & Electro-Optics Society (LEOS) Graduate Student Fellowship in 2004.

Michael Düser (S’99–M’03) received the Dipl.-Ing. degree in electrical engineering with distinction from the University of Dortmund, Dortmund, Germany, in 1998 and the Ph.D. degree in electronic and electrical engineering from the University of London, London, U.K., in 2003, where his research with the Optical Networks Group, University College London (UCL), London, U.K., focused on the theoretical and experimental investigation of advanced optical-packet-routed systems, networks, and associated devices, in particular wavelength-routed optical-burst-switched (WR-OBS) network architectures. He was formerly with the Optical Fiber Research Department of Bell Laboratories, Lucent Technologies, from November 1997 until October 1998, working on a research project on multimode optical fibers for high-bit-rate local area network applications. He has been a Research Fellow in the Department of Electronic and Electrical Engineering at UCL since April 2003, working on a collaborative project concerned with research of advanced quantum computation. Dr. Düser is a Member of the Institution of Electrical Engineers (IEE) and the German Verein Deutscher Elektrotechniker (VDE). He received an IEEE Lasers & Electro-Optics Society (LEOS) Graduate Student Fellowship in 2000.

2433

Jason Spencer (S’02) received the B.Eng. degree in electronic engineering and the M.Sc. degree in telecommunications from the University College London (UCL), London, U.K., in 1997 and 1998, respectively. He is currently working toward the Ph.D. degree at UCL on the interactions between network layers and their effects on network design. Most recently, he has been investigating interdomain Internet protocol quality-of-service architectures as part of the IST MESCAL project. His research interests include network planning and management, decentralized network control, next-generation high-speed reconfigurable networks, complex systems, and large-scale system design.

Polina Bayvel (S’87–M’89–SM’00) received the B.Sc. (Eng.) and Ph.D. degrees in electronic and electrical engineering from University College London (UCL), London, U.K., in 1986 and 1990, respectively. Her Ph.D. research focused on nonlinear fiber optics and their applications. She worked under a Royal Society postdoctoral exchange fellowship in the Fiber Optics Laboratory at the General Physics Institute in Moscow (USSR Academy of Sciences) in 1990. Subsequently, she worked as a Principal Systems Engineer at STC Submarine Systems, Ltd., Greenwich, U.K., and Nortel Networks (formerly BNR-Europe), Harlow, U.K., and Ottawa, Canada, on the design and planning of high-speed optical fiber transmission networks. In 1993, she was awarded a Royal Society University Research Fellowship, which she began in 1994 at UCL and held for ten years. In 2002, she was promoted to a Chair in Optical Communications and Networks at the Department of Electronic and Electrical Engineering, UCL. She currently heads the Optical Networks Group, focusing research on high-speed optical communication systems, fiber nonlinearities and transmission, new optical network architectures and algorithms, and associated devices. She has authored/coauthored more than 160 refereed journal and conference papers. Dr. Bayvel is a Fellow of the Royal Academy of Engineering (FREng), the Institute of Physics (FInstP), and the Institution of Electrical Engineers (IEE) and serves as Honorary Editor of its publication Electronics Letters. She is also on the TPC of the IEEE Lasers & Electro-Optics Society (LEOS) Annual Meeting and Optical Society of America (OSA) Conference on Lasers and Electro-Optics (CLEO). In 2002, she received the 2002 Paterson Prize and Medal from the Institute of Physics for her contributions to research on the fundamental aspects of nonlinear optics and their applications in optical communications systems. She is the Technical Chair for the European Conference on Optical Communications in 2005 (ECOC 2005), Glasgow, U.K.

Ignacio de Miguel (S’02–A’02–M’02) received the Telecommunication Engineer degree and the Ph.D. degree from the University of Valladolid, Valladolid, Spain, in 1997 and 2002, respectively. He has worked as a Junior Lecturer at the University of Valladolid since 1997. He has also been a Visiting Research Fellow with the Optical Networks Group, University College London (UCL), London, U.K. His research interests are the design and performance evaluation of optical networks, especially wavelength-routed optical networks and wavelengthrouted optical-burst-switched networks, as well as Internet protocol over wavelength-division multiplexing (IP over WDM). Dr. de Miguel received the Nortel Networks Prize for the Best Ph.D. Thesis on Optical Internet from the Spanish Institute and Association of Telecommunication Engineers (COIT/AEIT) in 2003. He also received the Innovation and Development Regional Prize for his Graduation Project in 1997.

2434

Dirk Breuer was born in Düren, Germany, in 1967. He received the Diploma degree in electrical engineering and the Ph.D. degree for studies on high-capacity optical transmission systems from the Technical University of Berlin, Berlin, Germany in 1993 and 1999, respectively. He was with Virtual Photonics, Inc., from 1998 to 2000, where he was responsible for developing a physical-layer design tool. In 2000, he joined the Photonic Systems Group at Technologiezentrum, T-Systems Nova GmbH, Deutsche Telekom, Berlin, Germany. The main focus of his research is currently on the optimization of the optical transport network of Deutsche Telekom and evaluation of new network technologies.

Norbert Hanik (M’00) was born in 1962. He received the Dipl.Ing. degree in electrical engineering (with a thesis on digital spread spectrum systems) and the Dr.Ing. degree (with a dissertation on nonlinear effects in optical signal transmission) from the Munich University of Technology, Munich, Germany, in 1989 and 1995, respectively. He was a Research Associate at the Institute for Telecommunications, Munich University of Technology, from 1989 to 1995, where he worked in the areas of mobile radio and optical communications. From 1995 to 2004, he was with Technologiezentrum, T-Systems Nova GmbH, Deutsche Telekom, Berlin, Germany, heading the research group System Concepts of Photonic Networks. During his work there, he contributed to a multitude of Telekom internal research and development projects, both as a Scientist and Project Leader. During 2002, he was a Visiting Professor at Research Center COM, Technical University of Denmark, Copenhagen. He is currently an Associate Professor of wired and optical signal transmission with the Munich University of Technology, Munich, Germany. He has participated in the German research projects KomNet and currently MultiTeraNet and the European Research Projects ACTS METON and DEMON. His primary research interests are in the fields of physical design, optimization, operation, and management of optical networks.

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 22, NO. 11, NOVEMBER 2004

Andreas Gladisch (M’00) was born in Germany in 1960. He received the Dipl.Ing. degree in theoretical electro techniques from the Technical University of Ilmenau, Ilmenau, Germany, in 1986 and the Ph.D. degree in optical communications from Humboldt University of Berlin, Germany, in 1990, where he was engaged in research on coherent optical communication and optical frequency control. He joined the Research Institute of Deutsche Telekom in 1991, where he was involved in projects with coherent optics, wavelength-division-multiplexing (WDM) systems, wavelength control, and frequency stabilization. From 1996 to 1998, he was Director of a research group working in the field of design and management of optical networks, and in 1999, he became the Director of the Department on Network Architecture and System Concepts of T-Systems Nova, Berlin, Germany, the research and development subsidiary of Deutsche Telekom, whose main responsibility is the development of migration concepts to introduce additional optical functionality into the networks and to represent Deutsche Telekom in the Optical Internet Forum (OIF) and other forums dealing with Internet protocol/WDM (IP-WDM) integration. He has participated in several European research projects, such as ACTS-Meton, ACTS-Demon, ACTS-Moon, and IST-Lion as well as EURESCOM P918, where his work focused on IP-WDM integration, functional network architecture of optical networks, and management requirements. He was involved in projects dealing with the short-term development of the Deutsche Telekom transport network, especially the migration of synchronous digital hierarchy (SDH) and WDM, including the development and assessment of different network scenarios. He has been responsible for a project covering the midterm strategy of the transport network and the harmonization of the transport networks of the Deutsch Telekom Group. Since 2004, he has been managing a project that is investigating the usage of next-generation optical systems and global multiprotocol label switching in the network of Deutsche Teledom. He has authored or coauthored more than 75 national and international technical conference or journal papers. Dr. Gladisch is a Member of Informationstechnische Gesellschaft (ITG).