Optical Flow Switching

Vincent W. S. Chan,

Fellow, IEEE, OSA,

Guy Weichenberg,

Student Member, IEEE,

and Muriel Médard

, Senior Member, IEEE

Laboratory for Information and Decision SystemsMassachusetts Institute of Technology(

Invited paper

)

Abstract

– In this work, we evaluate an attractive candidate for optical network data transport:Optical Flow Switching (OFS). We describe the operation and implementation of thearchitecture, characterize its capacity region and capacity-cost tradeoff, and compare it to otherprominent optical network architectures.

I. I

NTRODUCTION

In the early days of the Internet, the most preciousresource was long-haul transmission capacity. Thus,the packet switching architecture was designed to usethis resource as efficiently as possible (Figure 1). Withthe ubiquitous deployment of WDM in core networksin the last few years, long-haul transmission is nowcheaper than router switching at core nodes; this is because the rate of decrease in cost per unit of opticalfiber capacity is well below that of Moore’s Law,which governs the cost of electronic ICs from whichelectronic routers and switches are built. To keepreaping the economic benefits of optical networkingtechnologies, long duration (quasi-static) optical circuitswitching – GMPLS – will be used in the next cycle of build-up in the core. In access networks, electronicrouters and switches are currently used for aggregationof bursty computer communications data. With theinexorable increase in communications bandwidthdemand, such architectures will soon impede the trendof cost reductions, at which point optical access andsome form of optical switching and routing for largetransactions will be required. In this paper, we proposea transport architecture – Optical Flow Switching(OFS) – that will exploit optical switching, routing, andtransport technologies to continue the lowering of costsfaster than Moore’s Law. This will ultimately allowaccess of high-rate services to the masses sooner thancurrent trends would otherwise allow. II.

OFS

ARCHITECTURE

A. Overview of OFS

In OFS, users request end-to-end lightpaths for longduration (>100 msec) transactions (Figure 2). If possible, each transaction is scheduled to meet a timedelay requirement over a finite time horizon. If thelightpath request is granted, the resources dedicated

Figure 1. Network cost reduction as results of disruptive opticalnetwork technologies.

to the transaction are relinquished for other users oncethe transfer is complete. Since it may not be possibleto honor every lightpath request, there will be a trade-off among three network performance metrics: delay, blocking probability, and wavelength utilization. Thistrade-off lends to multiple service qualities coexistingin the same network, which may be implemented bysimply incorporating proper cost functions into thescheduling protocol.In order to schedule data transmission across theWAN, users communicate via an electronic control plane with the scheduling processors assigned to their respective MANs. These scheduling processors, inturn, coordinate transmission of data across the WANin an electronic control plane. In OFS, it is assumedthat the smallest granularity of bandwidth that can bereserved across the core is a wavelength. For implementation simplicity, we further assume thatwavelength conversion is not used in the network,although this assumption may be relaxed. In the eventthat several single users have transactions which arenot sufficiently large to warrant their own wavelengthchannels, they may multiplex their data for transmission across the WAN via dynamic broadcastgroup formation.

1970 1974 1978 1982 1986 1990 1994 1998 2002 2006 2010 2014 2018 2020

Fiber trunksIncreasing line speeds

Dispersion managede-switched architectureOptical switching

•

disruptive technology -WDM fiber links

•2

disruptive technology -optical switching

•

disruptive technology -direct optical access

•

disruptive technology -new transport mechanisms

Electronic accessOptical access

Limit of WDM/optical switching technology ?

Computing

1 1 0 1 0

1 0

Subscriber cost

1970 1974 1978 1982 1986 1990 1994 1998 2002 2006 2010 2014 2018 2020

Fiber trunksIncreasing line speeds

Dispersion managede-s

witched architectureOptical switching

•

disruptive technology -WDM fiber links

•2

disruptive technology -optical switching

•

disruptive technology -direct optical access

•

disruptive technology -new transport mechanisms

Electronic accessOptical access

Limit of WDM/optical switching technology ?

Computing

1 1 0 1 0

1 0

Subscriber cost

Figure 2. Transport mechanisms (EPS in blue, GMPLS ingreen, OFS is from user to user inred).

Motivated by the minimization of network management and switch complexity in the network core, flows are serviced as indivisible entities. That is,data cells comprising a flow traverse the network contiguously in time, along the same wavelengthchannel, and along the same spatial network path. Thisis in contrast to packet switched networks, wheretransactions are broken up into constituent cells, andthese cells are switched and routed through the network independently. Note that in OFS networks, unlike packet switched networks, all queuing of data occurs atthe end users, thereby obviating the need for bufferingin the network core. Core nodes are thus equipped with bufferless optical cross-connects (OXCs). OFS is acentralized transport architecture in that coordination isrequired for logical topology reconfiguration.However, OFS traffic in the core will likely beefficiently aggregated and sufficiently intense towarrant a quasi-static logical topology that changes oncoarse time scales. Hence, the centralized managementand control required for OFS is not expected to beonerous. The network management and control carriedout on finer time-scales will be distributed in nature inthat only the relevant ingress and egress accessnetworks will need to communicate.

B. Detailed description of OFS

The key to high utilization of backbone wavelengthchannels – the costly part of the network due to theneed for optical amplifiers and dispersion management – is statistical multiplexing of large flows from manyusers in a scheduled fashion. In a large access network,there is less aggregation closer to the user and hencethere is less cost sharing of expensive (usually active)optical components. Figure 3 shows a hierarchicalaccess network comprising a feeder network and manydistribution networks, passive optical LANs, fed by thefeeder network.Physical topology and other physical layer designWhereas many users share the feeder network,distribution networks will serve relatively few users.

Figure 3. Hierarchical optical access network.

Thus, motivated by cost, one may reasonably take theapproach of allowing costly active components in thefeeder network, but avoiding these components in thedistribution networks. Figure 4 shows a feeder network with active access nodes and bus/tree typedistribution networks with passive optical components.

Figure 4. Physical architecture of access network.

The feeder network itself can be a ring or mesh. Ithas active optical and electrical components andsupports lightpath reconfigurability and dynamicwavelength broadcast group formation in millisecondtime frames. Since there is already significant trafficaggregation in the feeder, important core network technology building blocks such as OXCs, tunablefilters, optical amplifiers, and add-drop multiplexersmay economically be used.The distribution network serves as the interface between the network and its users. Since fiber at theend user is less of a shared medium, bandwidth is ingreat abundance and may therefore be prudently tradedfor a low cost distribution network. For the physicaltopology of the distribution network, all-optical ringsare not viable since feedback generates oscillations.The topologies that can be used are buses and trees(and stars which are degenerate trees). The number of

User 1User 2Router 1Router 2Router 3

WDM layer

. . .. . .Network control

User 1User 2Router 1Router 2Router 3

WDM layer

. . .. . .Network control

COANANANANAN

Distribution Rings

Access Node

Distribution Tree

AccessNetwork BackboneNetwork

Distribution bus

Feeder network Ring or mesh

GatewayWAN

Feeder Network

Active components,monitoring, management,control, high utilization

Distribution Network

Passive components

Distribution Network

Passive components

Distribution Network

Passive components

Distribution Network

Passive components

Distribution Network

Passive components

users supported by a single distribution network should be the largest number possible without activeamplification in this part of the network. Passive busesand trees, however, impose severe limits on the number of users that can be supported by these distributionnetworks. Furthermore, fiber nonlinearities impose anupper limit on the amount of optical power that can belaunched into a distribution network from the accessnode. The number of users at ~2.5 Gbps that can besupported by a passive optical network is theoretically between 10 and 800, but realistically, the range will beapproximately 20-100. In some applications, such aslarge computer file transfer, users can be active withless than 1% duty cycle when they are in session. If only 10% of users on the network are in session at anygiven time, then to gain some degree of efficientstatistical multiplexing in a wavelength channel in afiber, the number of users accessing the wavelengthchannel should be on the order of 1000. Thus, we needan access network architecture that is better than that of conventional passive optical networks.Remotely-pumped EDFAs offer the possibility of greatly increasing the number of users supportable.The pump sources can be maintained at an access nodeon the feeder network simplifying deployment,management and control, and also conforming to thedesire to have no component in the distributionnetwork that requires electrical supply. In the design of such a distribution network, sections of erbium-dopedfiber are inserted at suitable locations along the lengthof undoped single-mode fiber, as shown in Figure 5.The architecture consists of a bus network with a chainof remotely pumped EDFAs. A tributary passive star or tree network is fed by a tap after the gain stage.This has the additional advantage that the EDFAs areconcentrated near the access node, and the shorter fiber length to them helps reduce absorption of the pump power by the fiber. Towards the middle to the end runof the bus when amplified spontaneous emission (ASE)is the dominant noise, the tapped power must increase(linearly) with the tap number. This is why theamplification process cannot be used to extend therange indefinitely. However, the maximum number of users supportable is over 10,000 and thus more thanadequately meets the aforementioned requirement of ~1000 users for efficient statistical multiplexing.Dynamically arranged broadcast groups within theaccess network are an additional key feature of our envisioned network architecture. They allow for arbitrary connectivity of users, efficient sharing of wavelength channels over time, flexibility in setting upOFS connections, and ease of multicast and broadcastservices. Dynamic broadcast groups can be formed viaassignment of specific wavelengths to a group of users.Access to these wavelengths is supported in thedistribution network via broadcast, but switching andwavelength conversion in the feeder network can beused to form the group. To further localize power distribution for high-rate applications and a high degreeof aggregation, remotely optically switched MEMs can be used to channel wavelengths in passive tributaries(Figure 6).

Figure 5. Remotely-pumped distribution network feeding end-users, passive stars, and trees.Figure 6. Remotely optically switched access network.

Higher layer designFigure 7 shows how end-to-end, bursty, all-opticalflows can be routed from the distribution network across a WAN to a peer access network. As discussed previously, access nodes can be fully reconfigurable,with programmable ADMs and OXCs.

DLogicMEM-sw

Optically powered switch

Switchingsignal

Switch

DLogicMEM-sw

Optically powered switch

DLogicMEM-sw

Optically powered switch

Switchingsignal

Switch

Pump source