OPTICAL PACKET SWITCHING NETWORKS
`
`The Application of
`Optical Packet Switching in
`Future Communication Networks
`
`Mike J. O’ Mahony,1 Dimitra Simeonidou,1 David K. Hunter,2 and Anna Tzanakaki
`Ilotron Engineering Centre
`
`ABSTRACT
`Telecommunication networks are experienc-
`ing a dramatic increase in demand for capacity,
`much of it related to the exponential takeup of
`the Internet and associated services. To sup-
`port this demand economically, transport net-
`works are evolving to provide a reconfigurable
`optical layer which, with optical cross-connects,
`will realize a high-bandwidth flexible core. As
`well as providing large capacity, this new layer
`will be required to support new services such
`as rapid provisioning of an end-to-end connec-
`tion under customer control. The first phase of
`network evolution, therefore, will provide a cir-
`cuit-switched optical layer characterized by
`high capacity and fast circuit provisioning. In
`the longer term, it is currently envisaged that
`the bandwidth efficiency associated with opti-
`cal packet switching (a transport technology
`that matches the bursty nature of multimedia
`traffic) will be required to ensure economic use
`of network resources. This article considers
`possible network application scenarios for opti-
`cal packet switching. In particular, it focuses
`on the concept of an optical packet router as
`an edge network device, functioning as an
`interface between the electronic and optical
`domains. In this application it can provide a
`scalable and efficient IP traffic aggregator that
`may provide greater flexibility and efficiency
`than an electronic terabit router with reduced
`cost. The discussion considers the main techni-
`cal issues relating to the concept and its imple-
`mentation.
`
`INTRODUCTION
`The rapidly increasing bandwidth demand, driv-
`en by the Internet, has led to a paradigm shift
`in the telecommunications industry from voice-
`optimized to IP-centric networks. In this sce-
`nario, the role of synchronous digital
`
`hierarchy/optical network (SDH/SONET) will
`diminish, and the optical transport network will
`directly provide a global transport infrastructure
`for legacy and new IP services. The utilization
`of optical networking employing dense wave-
`length-division multiplexing (DWDM) in con-
`junction with optical cross-connects (OXCs)
`presents many new opportunities for supporting
`faster and more flexible provision of legacy and
`IP services. A major driver for realizing this
`evolution is the potential ability of such net-
`works to provide fast automatic setup and tear-
`down of paths across the optical network, with
`the capability of supporting diverse client signals
`on the paths. The main focus, therefore, of
`today’s optical network planning lies in imple-
`menting a dynamically reconfigurable optical
`transport layer based on fast OXCs coupled
`with a suitable control and management archi-
`tecture [1]. Thus, in the near future an optical
`transport network (OTN) will be realized capa-
`ble of supporting large numbers of high-capacity
`optical channels, with bit rates on the order of
`10–40 Gb/s. This model of the network is illus-
`trated in Fig. 1. The diagram presents a possible
`OTN structure, which comprises the intercon-
`nection of a number of OXCs in a mesh topolo-
`gy. Since each interconnecting fiber may support
`many wavelengths (e.g., > 100) and there may
`be many fibers (e.g., 32), the OXCs require the
`capability to support the cross-connection of
`many thousands of wavelength channels. This
`OTN, therefore, will provide wavelength paths
`to clients such as IP routers, SONET/SDH net-
`work elements, and ATM switches, and Fig. 1
`illustrates how the network might interconnect
`two IP routers. In addition to the switching
`hardware a control layer is necessary to set up
`the network path, and this normally interacts
`with the OXC controller to initiate switching
`within the OXC. A signaling channel between
`nodes ensures that each OXC has knowledge of
`the network resource status, paths available, and
`
`1 University of Essex
`
`2 University of Strathclyde
`
`128
`
`IEEE Communications Magazine • March 2001
`0163-6804/01/$10.00 © 2001 IEEE
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 1 of 8
`
`

`

`Client
`
`so on. Current research focuses on the use of
`distributed management schemes such as multi-
`protocol label switching (MPLS) to provide the
`control plane necessary to ensure fast path
`setup. In this type of application the label is the
`wavelength of the incoming signal; hence, the
`term multiprotocol lambda switching is more
`commonly used.
`This dynamically reconfigurable optical
`transport network, therefore, will enable the
`fast allocation of high-capacity paths to link
`clients. Also, the pace of development is such
`that the technology (of transmission and switch-
`ing) will support huge numbers of optical chan-
`nels (wavelengths). It might therefore seem that
`in this future network bandwidth is not an
`issue, and optical circuit switching (the tech-
`nique we have been discussing) will meet all
`future requirements. However, this is not the
`case for a number of reasons. The OTN, for
`example, offers granularity only at the wave-
`length level to clients. Thus, if the traffic source
`is bursty, the channel capacity may be under-
`used, which will have an impact on the dimen-
`sioning of the network and the size of the
`OXCs. This argument is particularly strong as
`the network moves to become data- rather than
`voice-centric. Economics will always demand
`that the network resources be used efficiently.
`A major advantage of electronic packet switch-
`ing is its bandwidth efficiency and ability to
`support diverse services; hence, research is now
`focusing on bringing the packet switching con-
`cept into the optical domain, that is, optical
`packet switching (OPS).
`In this article the use of OPS in the future
`network is discussed. First a general look at its
`application areas is considered, however it is
`believed that the first application will be as an
`edge router interfacing the electronic IP domain
`and the OTN. In the succeeding sections the
`technical and implementation issues relating to
`this concept are discussed.
`
`Client
`
`Optical transport network
`
`IP router
`(or SONET/SDH)
`
`IP router
`
`OXC
`
`Controller
`
`I Figure 1. An optical transport network.
`
`in an all-optical manner, is still many years away.
`For medium-term network scenarios, OPS using
`electronic control and header processing is more
`realistic; indeed, it is not clear what major advan-
`tages the all-optical approach has to offer over
`this opto-electronic approach. This article focus-
`es on the approach in which the optical packet
`comprises an optical label (often realized using
`sub-carrier modulation techniques) attached to a
`payload, which may be of fixed or variable dura-
`tion (other approaches, e.g., burst or flow switch-
`ing, are not considered here). The client signal,
`such as IP packets, forms the payload, and the
`optical packet entity is routed through the net-
`work. Within an OPS, the packet header or label
`is read and compared with a lookup table. The
`payload is then routed to the appropriate output
`port with a new label attached (label swapping).
`An important feature is that the payload is trans-
`parently routed through the switch (i.e., stays
`within the optical domain), but the label pro-
`cessing and switch control are electronic. Some
`of these issues are discussed in more detail
`below.
`
`OPTICAL PACKET SWITCHING
`Research into OPS has been conducted over a
`number of years [2–4]. Pure OPS, in which pack-
`et header recognition and control are achieved
`
`APPLICATIONS
`The attractive feature of OPS is that it can
`appear as a natural evolution of the OTN. In
`particular, the OXCs developed for the OTN
`
`Label
`switch routers
`
`Edge
`OPS
`
`OXC
`
`Optical transport network
`
`SONET/
`SDH
`
`Core
`OPS
`
`SONET/
`SDH
`
`Edge
`OPS
`
`I Figure 2. Applications of OPS as core and edge routers.
`
`IEEE Communications Magazine • March 2001
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`129
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 2 of 8
`
`

`

`The first step
`toward optical
`data networking
`is the
`implementation
`of a network
`control plane,
`based on
`distributed label
`switched
`management
`principles such as
`the MPLS control
`model and
`associated with
`the OXC.
`
`can support an OPS network layer. Figure 2
`illustrates a network comprising OXC and
`OPS elements. As shown, resources can be
`used in a number of ways. For example, some
`optical channels (wavelength paths) may inter-
`connect high-capacity points that will fully uti-
`lize channel capacity, such as SDH rings.
`Other channels might be used to support opti-
`cal packet transmission for efficient use of
`bandwidth, to either optimize resource utiliza-
`tion within the network or, for example, sup-
`port an end-to-end point and click provisioning
`service where granularity may be an issue. Fig-
`ure 2 therefore illustrates two key OPS appli-
`cation scenarios. One is the application as a
`core switch. Packets traveling through the net-
`work undergo switching at core nodes where
`ongoing route selection and label swapping
`take place. In this mode OPS maximizes uti-
`lization of the network resources, minimizing
`the total network capacity required, reducing
`the size of the OXCs. The second application,
`the main source of discussion in this article, is
`that of an edge router interfacing the electron-
`ic IP domain to the OTN. This is illustrated in
`Fig. 2, which shows the packet switch posi-
`tioned as an edge router interfacing to both
`the OTN and IP domains. In this application
`the OPS provides a number of key functions
`required of the future OTN, as highlighted
`below.
`
`CONTROL PLANES
`As discussed above, the first generation of
`OXCs will not perform packet-level process-
`ing. The entire traffic on any optical channel
`at an input port in an OXC is switched to an
`output port; thus, the optical channel supports
`continuous data. IP traffic, however, cannot be
`constructed as continuous data streams; there-
`fore, there is a pressing requirement to devel-
`op the framework for a data/IP-aware optical
`transport.
`At present, data/IP services are provided
`through networks that may include three or four
`different electronic multiplexing and switching
`layers (e.g., IP, frame relay, ATM, SONET).
`The multiplicity of layers produces inefficiencies,
`adds to the latencies of connections, and inhibits
`the provisioning of quality of service assurances.
`Worse, the layers are largely unaware of each
`other, causing duplication of transport protocols
`and management tasks.
`The first step toward optical data networking
`is the implementation of a network control
`plane, based on distributed label-switched man-
`agement principles such as the multiprotocol
`label switching (MPLS) control model and asso-
`ciated with the OXC. The functions of this con-
`trol plane will initially be to establish and
`maintain optical paths within the network and,
`in the long term, to determine, distribute, and
`maintain state information associated with the
`OTN. This control plane will also be responsible
`for updating the information in the local switch
`controller (Fig. 1). As a result, the OXCs within
`the OTN will switch optical channels, in a simi-
`lar way to label switched routers (LSRs) switch
`packets in an electronic IP network. LSRs per-
`
`form packet-level operations using information
`carried on the labels attached to the data pack-
`ets, while with OXCs the switching information
`is inferred from the wavelength (MPlS) or opti-
`cal channel overhead.
`In networking systems involving a number of
`data clients and OXCs, MPLS can provide a uni-
`form control plane strategy in order to reduce
`the complexity of managing dissimilar network-
`ing systems. In these future network scenarios
`the question of where the boundary between the
`service and transport layer lies is still unan-
`swered, but there is clearly a need to maintain
`topology and control isolation as well as to cre-
`ate an efficient interface between the optical
`transport and the service layers.
`There are many reasons to separate network
`topologies and control, whether it is physical or
`logical for the OTN and the service layer. Some
`of the reasons are due to a number of important
`differences between electronic data routers and
`optical wavelength routers that necessitate spe-
`cial features to be implemented in the control
`plane. The first difference would be the band-
`width granularity, which is much coarser for an
`OXC than that for an IP router (wavelengths
`rather than packets). Because of the high-band-
`width nature of an optical connection, one would
`expect them to persist for a much longer dura-
`tion and involve relatively infrequent connection
`requests when compared to per-packet routing
`operations. A further specific requirement for
`the control plane will be to maintain OTN infra-
`structure information in order to facilitate path
`selection for optical channels. Information will
`include fiber characteristics, amplifier positions,
`and signal evaluation data. This information can
`be collected through optical supervisory channels
`and optical channel overhead processing, and
`can be actively used for setting up optical paths
`and fault localization.
`The most important reason, perhaps, for iso-
`lating the two layers is that they are likely to be
`under different administrative controls (or own-
`ership) and policies. Under such circumstances
`the service provider who owns the OTN will
`wish to maintain full control of the network.
`Such an operator would not wish to give a client
`insight into the structure of the OTN layer since
`this is his/her business value.
`Although the service provider does not wish
`to give the client knowledge of the OTN, there
`are client services that depend on having a view
`of the internal structure of the OTN. Three
`examples are suggested. The first involves con-
`nections diversely routed for provisioning and
`restoration purposes. The second involves a con-
`nection required at a future time, while the third
`involves being able to know which nodes can be
`reached via the OTN. Thus, network manage-
`ment features are required that allow limited
`internal OTN information to be accessed or
`manipulated by the client service layer in a man-
`ner that does not compromise the security of the
`operator’s network.
`Currently there are no router solutions that
`can satisfy all the above points and fit in a realis-
`tic future network solution able to carry effi-
`ciently a mixture of circuit- and packet-switched
`traffic into the OTN.
`
`130
`
`IEEE Communications Magazine • March 2001
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 3 of 8
`
`

`

`Network
`management
`system
`
`Node control
`
`1
`
`Output
`
`N
`
`Conditioning
`
`QXC
`
`Monitoring
`
`Optical
`packet
`switch
`
`SDH/SONET
`
`Clients (e.g., IP, ATM, MPLS)
`
`I Figure 3. Interfacing of the OPS with the OXC.
`
`from a number of sources and map onto optical
`packets. These optical packets will be of variable
`length, which will be an integral multiple of a
`chosen time unit. The aggregating nodes will
`then map the optical packets onto appropriate
`wavelengths for transport over the OTN to de-
`aggregating nodes that can either be egress
`points from the network or intermediary nodes
`that further map the optical packets onto new
`wavelength paths. During this process, the OPS
`will run a protocol capable of discovering the
`OXC network topology, and thus will be able to
`combine aggregation with QoS provisioning
`within the OTN.
`The optical router proposed here will provide
`a more scalable and efficient IP traffic aggrega-
`tor compared with similar electronic Terabit
`router solutions. Furthermore, it will take full
`advantage of the capacity, scalability and func-
`tionality provided by the optical layer, a function
`that cannot be provided by an electronic router
`solution.
`
`REALIZATION ISSUES
`The optical packet router will switch and buffer
`entities that may comprise multiple or single
`datagrams, or indeed only a part of one. To find
`the overall optimum packet transport solution
`for the optical edge aggregator, a number of
`issues need consideration, as described below.
`THE OPTICAL PACKET
`In order to reduce the number of entities that
`the switch must process per unit time, single or
`
`THE OPTICAL PACKET SWITCH AS
`EDGE ROUTER AND AGGREGATOR
`To overcome these limitations, an OPS solution
`can be used to facilitate efficient provisioning of
`packet services through the predominantly cir-
`cuit-switched OTN infrastructure. The OPS will
`fit in a realistic network scenario where circuit-
`and packet-switched traffic will be transported
`together through the OTN. The optical packet
`switching functionality will then coexist with
`wavelength routing provided through the OXCs.
`In this case, fast switching will be provided for
`the packet traffic where granularity below the
`wavelength level is required, while slow wave-
`length switching and routing will be facilitated at
`the same time. Fast switching and packet traffic
`aggregation for efficient bandwidth utilization
`will mainly be performed at the edge of the net-
`work (the interface with the IP/ATM domain)
`where dynamic and fast wavelength allocation
`for packet traffic will be required. Under this
`scenario, the OPS router will be an edge net-
`work device, which will function as a topological
`and logical interface between the service and
`transport layers. The OPS router can directly
`interface with the OXC, which will make a set of
`static wavelength and fiber routes available to
`the OPS traffic. This is illustrated in Fig. 3,
`which shows an OXC making up a central switch
`fabric capable of interconnecting the demulti-
`plexed input wavelength channels to the appro-
`priate outgoing fibers. Interconnection is
`controlled through the management and control
`subsystems. The OPS is positioned in the add-
`drop ports of the OXC and accesses wavelength
`channels dedicated to packet switching.
`The external electronic routers and OPS will
`handle the same granularity (per packet), which
`will lead to an integrated control plane between
`the IP and the OPS domains. At the same time,
`the OPS will maintain information on the config-
`uration, the physical infrastructure, the topology
`and scale of the OXC transport. Therefore, the
`proposed OPS will be able to isolate the OTN
`from the service layer while interfacing fully with
`both layers:
`• With the data/IP domain through integrated
`management control
`• With the OTN by maintaining information
`on the configuration, the physical infra-
`structure, the topology and scale of the
`OXC transport
`An additional benefit of the OPS will be due
`to the increased granularity over pure DWDM
`networks, which permits more efficient use to be
`made of the core network. One of the main dis-
`advantages of an OTN is that there is currently
`no mechanism to provide direct access to the
`OTN with bandwidth granularity that is finer
`than a whole wavelength. Providing this finer
`granularity is central to creating a network that
`is efficient, from the perspective of the operator,
`and cost effective, for the operator’s customer.
`A schematic diagram of the OPS functionality
`as an edge aggregator/router is presented in Fig.
`4. Here the OPS will provide an aggregation
`mechanism in the external OTN nodes that can
`accept packet type transport (i.e. IP and ATM)
`
`IEEE Communications Magazine • March 2001
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`131
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 4 of 8
`
`

`

`Optical packets are sent over OTN
`wavelengths. Contention resolution is
`based on QoS class implied from the
`label on optical packet.
`
`Packets are aggregated based on
`destination and QoS parameters, and
`then formed into optical packets with
`labels that signify destination and QoS
`class. (Diagram signifies two
`destinations with two QoS classes for
`one of the destinations, giving three
`label values).
`
`Ingress to node (and network) from
`multiple sources to multiple
`destinations (shown as two header
`shadings on packets).
`
`OTN1
`
`OTN2
`
`OPS
`
`packets
`Optical
`
`op1
`
`op2
`
`op3
`
`Packet format adaptation, memory
`classification
`
`s1
`
`s2
`
`s3
`
`s4
`
`s5
`
`Service layer
`
`I Figure 4. OPS functionality as an edge aggregator/router.
`
`multiple packets with the same destination and
`quality of service (QoS) class may be grouped
`together forming an optical packet at the edge
`of the network. The optical packet will be of
`variable length, which will be integral multiple of
`a unit length. While this reduces the complexity
`of the packet switches, it increases the complexi-
`ty of the interface at the edge of the network, in
`fact the complexity of forming the optical packet
`is comparable to implementing some of the
`functions of an IP router. Great care must be
`taken when designing a packet scheduling algo-
`rithm for this type of switch, to ensure that the
`algorithm can be implemented in real-time by
`electronic control hardware. The optical packet
`switched network now looks very much more
`like a burst switched network [5], the major dif-
`ference being that control information is still in-
`band. The implementation of the optical packet
`can be advantageous for the edge router applica-
`tion, where the optical router will perform and
`replace some or all of the Terabit router func-
`tionality.
`With this approach, the header must also be
`read, and the label must be translated electron-
`ically, in the usual way. In an MPLS-based
`approach such as that considered here, the
`header translation hardware will search in a
`table for the label held in the packet header.
`The entry in the table for that particular label
`will contain the new label (which must then
`over-write the existing label in the header), and
`the output to which the packet must be for-
`warded. Label stacking is very difficult to imple-
`ment in such an OPS since such an operation
`effectively involves changing the length of the
`header.
`
`BUFFER MEMORY IMPLEMENTATION
`
`To preserve an all-optical data path, it would be
`desirable to implement the buffer memory in the
`OPS optically. However, optical memory is in a
`relatively primitive state; there is no such thing as
`optical random access memory (RAM), and it is
`necessary to resort to fiber delay lines for memo-
`ry. If these become unduly long (tens or hundreds
`of kilometers), they become very costly, bulky,
`and difficult to stabilize with respect to tempera-
`ture. Here, a compromise is proposed where elec-
`tronics and optics share the buffering. Optics is
`used for very short delays, which form the vast
`majority of storage, and electronics is used for
`longer delays. The amount of electronic memory,
`with its costly electrical-to-optical and optical-to-
`electrical interfaces, is reduced. If a packet must
`be delayed more than the longest optical delay, it
`is passed to the electronic memory.
`This approach is particularly suitable for the
`edge aggregator because of the ability to make
`use of the electronic memory already in place to
`perform the electronic router functionality. In
`this case, although the switch will be able to
`operate without using the optical delay line
`memories, it can be shown that the optical mem-
`ories can be used for the majority of the buffer-
`ing (i.e., low delays) while only a fraction of the
`electronic memory is used for the higher delays,
`representing the minority of buffering.
`Initial simulation studies have demonstrated
`the validity of this technique, as illustrated in
`Fig. 5a, which shows the probability that a ran-
`domly chosen byte stored within an output-
`buffered packet switch is experiencing a delay
`greater than a given value within the switch. In
`
`The optical
`packet router will
`switch and buffer
`entities that may
`comprise multiple
`or single
`datagrams, or
`indeed only a
`part of one.
`To find the
`overall optimum
`packet transport
`solution for the
`optical edge
`aggregator, a
`number of
`issues need
`consideration.
`
`132
`
`IEEE Communications Magazine • March 2001
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 5 of 8
`
`

`

`1.00E+00
`
`1.00E-01
`
`1.00E-02
`
`1.00E-03
`
`1.00E-04
`
`1.00E-05
`
`1.00E-06
`
`Probability
`
`0
`
`5000
`
`10,000
`
`25,000
`
`30000,
`
`15,000
`20,000
`Delay (bytes)
`
`(a)
`
`2 m
`
`3 m
`
`2 m
`
`1 m
`
`(b)
`
`(c)
`
`I Figure 5. a) The probability that a randomly chosen byte is experiencing a
`delay greater than the given value in an output-buffered switch, for self-similar
`traffic, with a mean traffic level of 80 percent; b and c) an example of sharing
`delay lines. The subsystem in b) requires 5 m of fiber delay line, whereas only
`3 m are required in c), because the smallest delay line has been moved in front
`of the splitter.
`
`SWITCH IMPLEMENTATION
`A generic structure of the proposed packet
`switch consists of an input processing interface, a
`switching and buffering block, and an output
`processing module, as illustrated in Fig. 6a. The
`input interface performs delineation (i.e., identi-
`fication of the packet start and end), packet for-
`mat adaptation into the optical packet,
`classification into forward equivalent classes
`defined for the OTN, and electronic buffering.
`The switching and buffering blocks are responsi-
`ble for the routing of the optical packets to the
`appropriate output ports and contention resolu-
`tion, respectively, while the output interface is
`responsible for header reinsertion and per pack-
`et conditioning such as wavelength conversion to
`the appropriate OTN wavelengths, regeneration,
`and power equalization. The proposed architec-
`ture is based on a feedback buffering scheme to
`enable maximum utilization and sharing of the
`available buffers. The recirculating buffers used
`in this architecture offer the ability to support
`QoS classes via packet preemption. Header
`detection and processing are performed in the
`electronic domain.
`Fast optical switching per packet can be per-
`formed using a switch matrix based on semicon-
`ductor optical amplifier (SOA) gates or
`electro-optic technology. However, both tech-
`nologies are scalable only up to a limited switch
`
`an effort to mimic real Internet traffic, each out-
`put buffer is fed with self-similar traffic multi-
`plexed from 404 sources with truncated Pareto
`on/off periods. Each burst coming from a source
`consists of packets generated according to a
`packet length distribution obtained from real
`traffic. The alpha parameter is 1.1, and the mean
`total load per output is 80 percent. For this type
`of traffic, these results apply for any output
`buffer experiencing a load of 80 percent, on any
`size of switch.
`Suppose that optical delay line buffering han-
`dles all optical delays of 6750 bytes or less (i.e., a
`maximum fiber length of 1.08 km at 10 Gb/s);
`then, on average, only 10 percent of the bytes in
`memory at any time are experiencing a larger
`delay, and are buffered in electronic memory. If
`the maximum optical delay is increased to 11,500
`bytes, this corresponds to a maximum fiber
`length of 1.84 km and only 1 percent of bytes in
`electronic memory. Hence there is a trade-off
`between electronic and optical memory use,
`which can lead to a significant reduction in elec-
`tronic memory by employing only short lengths
`of shared delay line optical buffers.
`
`SCHEDULING AND CONTROL
`Since optical memory is implemented with delay
`lines and not RAM, the electronic scheduler for
`the architecture must direct the packets over the
`correct delay lines to make the architecture per-
`form the same function as one constructed from
`RAM buffers. The packet scheduling algorithms
`for the transport solutions discussed above can
`be implemented using high-speed electronics,
`and must consider issues such as fairness, imple-
`mentation of QoS classes, queue stability, and
`queue starvation. The trade-off between elec-
`tronic and optical buffering must be determined
`based on cost considerations. Initial results
`demonstrate that using an OPS to interface with
`electronic routers can produce cost savings in
`the network.
`Due to the bulkiness and expense of large
`amounts of delay line fiber, two techniques may
`be used to reduce the total length of fiber
`required; this impacts upon the control algorithm:
`• Multiple packets on different wavelengths
`may pass along a specific delay path simul-
`taneously. For example, the total length of
`fiber delay line required could be reduced
`by a factor of 16 by using one delay line
`with 16 wavelengths instead of 16 delay
`lines of equal length; in both cases the same
`number of wavelength converters are
`required (i.e., 16).
`• The total length of fiber delay line memory
`can also be reduced by sharing fibers between
`different delay paths — a simple instance of
`this technique is shown in Fig. 5b, c [6]. This
`principle can be extended so that a large
`array of fiber delays can be replaced by mul-
`tiple delay line stages, with a dramatic reduc-
`tion in the amount of fiber required.
`Packet scheduling algorithms should be
`amenable to parallel implementation in order to
`enable implementation on programmable gate
`arrays to run in real time. Also, the implementa-
`tion must scale to large switches such as will be
`encountered in practice in future.
`
`IEEE Communications Magazine • March 2001
`Authorized licensed use limited to: Riva Laughlin. Downloaded on April 20,2023 at 16:02:18 UTC from IEEE Xplore. Restrictions apply.
`
`133
`
`Ex. 1026
`CISCO SYSTEMS, INC. / Page 6 of 8
`
`

`

`System
`Back-to-back
`
`10
`20
`30
`40
`50
`Number of cascaded switches
`
`60
`
`(b)
`
`20
`
`18
`
`16
`
`14
`
`q factor (dB)
`
`12
`
`0
`
`1
`
`N
`
`Buffering
`
`Switching
`
`1
`
`N
`
`Input
`processing
`
`Output
`processing
`
`Electronic control
`
`(a)
`
`I Figure 6. a) A schematic diagram of the generic structure of the optical packet switch; b) concatenation performance of the wave-
`length converter and AWG arrangement; back-to-back measurement shown for comparison.
`
`dimension and require some form of synchro-
`nization at the input of the switch matrix. An
`alternative solution that enables fast transparent
`switching of individual packets enabling asyn-
`chronous operation of the switch matrix is based
`on tunable wavelength converters followed by a
`wavelength routing device such as an arrayed
`waveguide grating (AWG). In this case, routing
`of the packets to the required output ports of
`the switch is performed by controlling the wave-
`length of the incoming packets through the input
`wavelength conversion stage and subsequent
`transmission through the AWG. Optical wave-
`length conversion is performed through SOA-
`based converters using either cross-gain
`modulation or cross-phase modulation tech-
`niques. Using either of the two schemes, a con-
`tinuous wave (CW) source is needed, and in the
`case of tunable wavelength conversion this
`source is required to be either a fast tunable
`laser or a switchable laser array. The tuning
`speed of the converter is then determined by the
`tuning speed of the CW signal, which can be as
`fast as a few nanoseconds; thus, the switching
`speed will also be in the nanosecond regime.
`The overall switch matrix scales with the dimen-
`sion of the AWG router, which currently can be
`as high as 128 x 128. This approach was evaluat-
`ed in project WASPNET [7, 8]. The concatena-
`tion performance of this configuration was
`evaluated through recirculating loop experiments
`[9], and Fig. 6b shows measured Q factor for
`both back-to-back and system (i.e., AWG and
`wavelength converter) configurations. The results
`demonstrate penalty-free operation for up to 25
`cascaded nodes.
`The buffering functionality is provided
`through a combination of electronic and optical
`buffering . The wavelength agility offered using
`wavelength conversion on a per packet basis
`enables statistical multiplexing at the fiber band-
`width capacity level. Tunable wavelength con-
`
`verters may significantly reduce the buffering
`requirements by appropriately wavelength trans-
`lating optical packets so that they can be stored
`within the same fiber delay line. This not only
`simplifies the buffering schemes, but also has the
`advantage of suppressed transfer delay and
`packet delay variation due to the reduction of
`the depth of the optical buffers [8].
`
`SUMMARY AND CONCLUSIONS
`This article presents a novel and efficient solu-
`tion for a fully data/IP (i.e., IP and ATM) aware
`optical transport network. The current proposal
`is to use an optical packet switching technology
`in order to:
`• Reduce the number of network layers, to
`simplify network management software and
`remove associated transport overheads
`• Offer efficient traffic agg