Optical switch fabrics for ultra-high-capacity IP routers...

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 11, NOVEMBER 2003 2839

Optical Switch Fabrics for Ultra-High-CapacityIP Routers

Jürgen Gripp, Member, IEEE, Marcus Duelk, Member, IEEE, John E. Simsarian, Ashish Bhardwaj, Member, IEEE,Pietro Bernasconi, Member, IEEE, Oldrich Laznicka, and Martin Zirngibl

Invited Paper

Abstract—Next-generation switches and routers may rely onoptical switch fabrics to overcome scalability problems that arisein sizing traditional electrical backplanes into the terabit regime.In this paper, we present and discuss several optical switch fabrictechnologies. We describe a promising approach based on arrayedwaveguide gratings and fast wavelength tuning and explain thechallenges with respect to technical and commercial viability. Fi-nally, we demonstrate an optical switch fabric capable of 1.2-Tb/sthroughput and show packet switching with four ports running at40 Gb/s each.

Index Terms—Array waveguide gratings, burst-mode receivers,optical packet switching, terabit routers, tunable lasers.

I. INTRODUCTION

I NTERNET traffic has been steadily increasing and futuregrowth may dramatically accelerate once content providers

and broadband users begin to utilize the full potential of theglobal network. Traffic has doubled every six months overthe last eight years, and data transport has kept pace with thisexplosive bandwidth demand thanks to the deployment ofdense wavelength-division multiplexing (DWDM) and steadilyincreasing line rates. A single fiber can now support severalterabits/second with individual line rates of up to 40 Gb/s.

Internet protocol (IP), multiprotocol label switching (MPLS),and asynchronous transfer mode (ATM) switches, and on theother hand, have been doubling capacity only every 12 monthsand could potentially become a bottleneck in the near future.Single-bay routers with capacities above 100 Gb/s are nowavailable, but the number of backplane interconnections andthe power density is reaching physical limits. Thus, high-endrouters are moving to optical interconnects between largeelectrical switch fabrics and line cards that are distributed overseveral bays of equipment. These multibay solutions occupyvaluable floor space, consume too much power for many centraloffices, and pose reliability concerns due to the large numberof active components in the switch fabric. They also imposehigh initial investments even for partially loaded configurations

Manuscript received February 12, 2003; revised June 20, 2003.J. Gripp, M. Duelk, J. E. Simsarian, A. Bhardwaj, P. Bernasconi, and M. Zirn-

gibl are with Bell Laboratories, Lucent Technologies, Crawford Hill, NJ 07733USA.

O. Laznicka is with Integrated Network Solutions, Lucent Technologies,Westford, MA 01886 USA.

Digital Object Identifier 10.1109/JLT.2003.819150

Fig. 1. Architecture of a packet router with an optical switch fabric.

since the fabric comprises a substantial part of the cost. Routerswith optical fabrics, especially those with passive fabric cores,can scale to higher capacities, increase reliability, and at thesame time significantly reduce footprint, power consumption,the initial cost of a partially equipped system, and the total costas well.

II. HIGH-CAPACITY PACKET SWITCHES

The basic architecture of a packet router with an opticalswitch fabric is shown in Fig. 1. Line cards connect to thenetwork on one side and via the fabric to each other. The fabricis controlled by a central scheduler that manages the data flowthrough the router. The only difference from an electrical routerlies in the use of an optical switch fabric instead of an electricalbackplane (or optical interconnects) and an electrical fabric.Since the difference is internal, the router is fully compatiblewith existing networks. This paper focuses on optical technolo-gies that can be used in this router architecture. Specifically, wedo not discuss optical packet switching for all-optical networks,including such topics as optical label swapping and opticalbuffering, since these areas of research require new standardsand network architectures.

Fig. 2 shows a diagram of a line card. The ingress part ofthe line card receives data from the network through a stan-dard interface, deframes the data into packets, and processesthe address information on a per packet basis. It then buffersthe packets so that they are ready to be sent to the switch fabricwhen the scheduler establishes a connection to the appropriateegress line card. Data packets may be fragmented or aggre-gated into data envelopes that match the rate at which the sched-uler reconfigures the fabric [1]. The egress side reassembles

0733-8724/03$17.00 © 2003 IEEE

2840 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 21, NO. 11, NOVEMBER 2003

Fig. 2. Block diagram of a line card. The switch fabric interface can bephysically part of the line card, but it functionally belongs to the optical switchfabric.

data envelopes into packets, which are framed and sent into thenetwork. The connection between the line card and the switchfabric interface, indicated in Fig. 2 by a dashed line, can poten-tially be a standard, which allows the switch fabric to be usedwith different types of line cards.

III. OPTICAL SWITCH FABRICS

Optical fabrics can be used for two main applications: circuitswitches and packet routers. Optical fabrics for circuit switchesrequire relatively slow switching speeds, on the order of mil-liseconds, but must potentially handle terabit/s or even petabit/scapacities when used for traffic management and bandwidthprovisioning in the core of the network. There are a numberof technologies available for circuit-switched networks, such asmicroelectromechanical systems (MEMS) [2], [3], total internalreflection in fluid-containing planar lightwave circuits [4], andthermooptical interferometric switches in silica or polymers [5].All of these techniques offer low power consumption and trans-parency with respect to bit rate and modulation format. In thecase of MEMS, extremely large switches with thousands ofports have been demonstrated and switch fabrics have been of-fered commercially [6].

Optical switch fabrics for packet routers, on the other hand,require switching times in the nanosecond regime. Several tech-nologies have been used to build active electrooptic switch fab-rics, e.g., LiNbO matrix switches [7], [8] and semiconductoroptical amplifier (SOA)-based matrix switches [9], [10]. Thesematrices consist of large numbers of optical 22 switches thatare connected in Banyan-type configurations. In the case ofLiNbO , the high insertion loss of the individual switch ele-ments has limited the total number of ports to eight or 32 so far.The SOA matrices are not yet a mature technology and 88fabrics have been the maximum size due to amplified sponta-neous noise crosstalk and heat dissipation. Besides the scala-bility problems, matrix switches have the disadvantage that thefailure of a single element disables the entire fabric.

A broadcast-and-select technique has been demonstrated inwhich each one of the input broadcasts its data via powersplitters to outputs, and each output selects an input using2 SOAs [11], [12]. The approach has the advantage of re-moving active components from the fabric core, but the largenumber of SOAs required at the output ports limits its scala-bility and increases the cost of the system.

Another recently developed switching technique involveselectrooptical beam deflection in combination with an opticalphasar [13]. This technique requires very complex electrical,mechanical, and optical engineering of the fabric, especially

Fig. 3. Basic architecture of the optical switch fabric. The transmitters and thereceivers are part of the switch fabric interface.

for large port counts, and the reliability and scalability are stillopen issues.

A switching technique based on arrayed waveguide gratingrouters and fast wavelength tunable lasers is the focus of thispaper. As we will describe in Sections IV–VIII, the approachemploys commercially available components, the number ofcomponents scales linearly with the size of the fabric, and thecore is passive. These are key features that enable the design ofa reliable optical switch fabric that is space and cost-efficient.

IV. A RCHITECTURE

Fig. 3 shows the architecture of the optical switch fabric. Atthe core is a passive, strictly nonblocking crossbar, realized byan arrayed waveguide grating. The arrayed-waveguidegrating (AWG) routes light based on its wavelength from anyinput port to any output port, similar to a diffraction grating.Data envelopes arriving from the ingress side of the line cardsare sent via wavelength tunable transmitters through the AWGinto burst-mode receivers and from there to the egress line cards.The scheduler configures the wavelengths of the transmittersand thereby the connection pattern between transmitters and re-ceivers.

A physical design advantage of this architecture lies in thefact that the transmitters and receivers are part of the switchfabric interface in Fig. 2. By distributing the active switch com-ponents and using a passive device for the fabric core, the failureof an active component does not affect the entire system, whichis an important reliability factor.

The wavelength-routing properties of the AWG are describedin Section V. The tunable transmitters are described in Sec-tion VI, the burst-mode receivers in Section VII, and a switchfabric demonstration in Section VIII.

V. ARRAYED WAVEGUIDE GRATINGS

Arrayed waveguide gratings play an important role intelecommunications networks. First proposed in form of aphased-array demultiplexer in 1988 [14], and subsequentlyextended from a 1 to an design [15], AWGs are nowwidely used in dense WDM systems. Today, silica-based AWG(de)multiplexers are available with standard specificationsof, e.g., 40 channels or more, 50-, 100-, or 200-GHz channelspacing and less than 4-dB insertion loss. Large AWGs[16] are also becoming commercially obtainable [17].

GRIPPet al.: OPTICAL SWITCH FABRICS FOR ULTRA-HIGH-CAPACITY IP ROUTERS 2841

Fig. 4. N �N arrayed waveguide grating.

AWGs are two-dimensional imaging systems withwavelength diffraction properties. As shown in Fig. 4, the roundsurfaces of the planar regions and the curvedwaveguides ofthe grating allow for a light source positioned on the object planeof the first planar region to be reconstructed at the image planeof the second planar region. The linearly increasing length of thegrating waveguides introduces a linear phase shift on the lightfront, which causes the location of the image to be wavelengthdependent. With a proper positioning of the inputs and outputs,light can travel from any input to any output port if it has the ap-propriate wavelength. Similar to conventional diffraction grat-ings, AWGs generate multiple images, one per diffraction order,and possess a periodic spectral response, known as the gratingfree spectral range (FSR). The design subtleties go beyond thescope of this paper and can be found elsewhere [16], [18]. InSectionV-A–C, we focus on the properties of these devices inthe context of wavelength switching.

A. Connection Matrices

AWGs can be designed so that the same set ofwavelengths is used to propagate light from any input to anyoutput port. This is achieved by choosing the FSR to be equal to

times the channel spacing . In that case, the connectivityis of a cyclic nature and the connection matrix has the form

......

.. ....

... (1)

where each element contains the wavelength that connectsinput port with output port . shows another consequenceof the inherent symmetry of the AWG. Each input connectsthrough different wavelengths to each one of theoutputsand in the same way, each output receives light ofdifferentwavelengths from the inputs.

As we will see in Section V-B, there are cases where thecyclic behavior is not the preferred choice. A second common

Fig. 5. Transmission spectrum of a noncyclic 42� 42 AWG. The channelspacing is 100 GHz; the Gaussian passbands have a FWHM of 50 GHz.

design is obtained by choosing FSR . The connec-tion matrix for this noncyclic case is given by

......

.. ....

... (2)

In this case each input requires a different set ofwavelengths,shifted by one channel from one input to the next. As a result,2 1 total wavelengths are required to obtain the connectivityof an crossbar.

B. Loss Nonuniformity

The loss nonuniformity of an AWG is primarily related to thefield distribution radiated by the grating array into the planar re-gion. Typically, the intensity envelope has a bell-like shape cen-tered in the middle of the image plane. Due to the energy distri-bution among the diffraction orders, in a cyclic AWG the outer-most output ports suffer almost 3 dB more loss than the centerone. An additional 3 dB loss occurs in moving from the centerinput port to an edge port, leaving us with a total loss nonunifor-mity close to 6 dB and the insertion loss typically ranging from2 to 8 dB. This could have potential negative consequences foran optical switch fabric by causing large connection-dependentpower variations between packets arriving at a receiver. For-tunately, these variations can be effectively eliminated by ad-justing the input power levels to inversely match the loss pro-file, i.e., by sending 3 dB less power into the center port thaninto an edge port. The maximum loss of 8 dB can, however, bean issue in system designs with tight optical power budgets. Inthis case, a noncyclic AWG is advantageous. With a much largerFSR, loss nonuniformity and worst case insertion loss can be re-duced, allowing for devices with larger port numbers.

Fig. 5 shows a spectrum of a noncyclic 4242 AWG. Theloss nonuniformity is better than 1.5 dB and the worst case in-sertion loss less than 4.2 dB in a fully packaged device.

C. Channel Spacing Nonuniformity

Due to the grating properties, it is easier to designAWGs with a uniform spacing in wavelength than in frequency.For instance, for a cyclic 32 32 AWG with 100-GHz channelspacing, some port combinations have a deviation from the fre-quency-based ITU grid of approximately12 GHz. More so-phisticated designs can reduce this to7 GHz [19] but nev-


ertheless the problem persists when the number of channelsgrows large. Wavelength misalignment between ITU-lasers andnon-ITU AWG passbands can cause additional loss and signaldistortions. One can restrict the AWG size such that the im-pairments are manageable, thus absorbing the penalties in thesystem design margins.

On the other hand, every laser could be recalibrated for theactual AWG passband positions instead of the ITU channels.This requires different sets of laser channels for every input portto the AWG, which can pose logistics problems in a commercialsystem. Nonetheless, such a step enables the use of even largerAWGs and can increase the total achievable throughput.

D. Passband Filtering and Crosstalk

The shape of the AWG passbands affects the system perfor-mance in two ways. First, if the sidebands of the modulateddata signal are filtered too strongly by a narrow passband, thedata will be distorted. The amount of distortion depends on theratio between passband width and data bit rate, the modula-tion format, and the exact shape of the passband. For example,sending a 40-Gb/s nonreturn-to-zero (NRZ) signal through apassband with Gaussian shape and 50-GHz full-width at half-maximum (FWHM) results in a roughly 1-dB penalty [20]. Thiscan be reduced with a flattop passband shape or a modulationformat with narrower sidebands, e.g., duobinary [21].

The penalties can also be decreased by using wider pass-bands, but this will lead to increased adjacent-channel crosstalk.For an AWG with 100-GHz channel spacing and 50-GHzFWHM Gaussian passbands, adjacent channel crosstalk istypically better than 25 dB, whereas the nonadjacent crosstalkcan be less than 35 dB. Out-of-band crosstalk has little effectat these levels, but in-band crosstalk between next-neighborscan cause system penalties. As a result, the system has to bedesigned with enough margin to handle worst case scenarios,such as all lasers transmitting at the same wavelength. It mayalso be possible to incorporate rules in the scheduling to avoidthe very small number of worst case connection patterns. Inthis case, the design specifications can be tightened and higherfabric throughput can be achieved by reducing the channelspacing and increasing the number of ports within the samewavelength range.

VI. FAST AND WIDELY TUNABLE LASERS

Wavelength tunability in lasers was achieved a few years afterthe invention of lasers themselves [22]. The use of tunable lasersin optical fabrics for packet switches, however, poses additionalrequirements that have only been met by fairly recent designs.The first requirement is tuning speed.

Today’s data traffic consists to a large extent (40%) of veryshort packets, typically 40 bytes in size [23]. At a bit rate of 40Gb/s, such a data packet is only 8 ns long. Multiterabit routerstypically have to aggregate these data packets into envelopeswith durations on the order of a microsecond to keep the work-load of the scheduler within acceptable limits [1]. As a conse-quence, the tuning speed of the lasers has to be in the tens-of-nanosecond range to keep the switching overhead reasonablysmall. This eliminates designs that are based on mechanical orthermal tuning. Electrically tunable semiconductor lasers, such

Fig. 6. (a) SG-DBR laser. The vertical arrows indicate injection currents.(b) GCSR laser. The white arrow shows how the light couples between twowaveguides with different index of refraction, assisted by a long period grating.

as distributed Bragg reflector (DBR) lasers, however, are wellsuited to meet this requirement. These devices can be wave-length-tuned by changing injection currents and thus varying theindex of refraction in one or more tuning sections. The wave-length tuning speed is limited by the effective carrier lifetime,which is typically a few nanoseconds.

The second requirement is wavelength tuning range. Opticalfabrics must provide high throughput, in the terabit/secondrange, to offer significant advantages over electrical fabrics.With the architecture shown in Fig. 3 and a cyclic AWG, therequired frequency tuning range is approximately thetotal throughput divided by the spectral efficiency

(3)

where is the bit rate, is the number of channels, andis the frequency spacing between channels. For a fabric with anoncyclic AWG, doubles unless lasers with different centerwavelengths are used at the different input ports [see (2)]. Fora total throughput of 1 Tb/s and a typical spectral efficiency of

bits/Hz, the required tuning range is about 2.5 THzor 20 nm (or even twice that), which exceeds the tuning rangeof conventional three-section DBR lasers. There are, however,a number of DBR laser designs with more than three sectionsthat provide tuning ranges up to 40 nm or 5 THz. These designsappear to be ideally suited for the switch fabric, and we brieflydescribe them in Section VI-A. A tuning range of 5 THz in com-bination with a modulation format that supports high spectralefficiency ( bits/Hz [21]) could lead to a total fabricthroughput of 4 Tb/s. Fabrics with throughputs in the tens ofterabits/second or beyond could be achieved by modifying thearchitecture of Section IV to include parallel data planes or mul-tiple stages.

A. Laser Designs

Fig. 6 schematically shows the structures of two designs:(a) the sampled-grating (SG) DBR laser and (b) the verticalgrating-assisted codirectional coupler with rear sampled gratingreflector (GCSR) laser. Both have one gain section and three


Fig. 7. Block diagram of the module electronics.

tuning sections. In the case of the SG-DBR laser [24], two sam-pled gratings whose reflection spectra have different peak spac-ings allow for broad tuning due to a Vernier effect. The GCSRlaser achieves broad tuning with a coupler region [25] that actsas a bandpass filter that selects one of the reflection peaks of asampled grating. The lasers have a phase section that providesfine tuning. Both designs achieve a tuning range of 40 nm andoutput power above 0 dBm. The wavelength is controlled by ac-curate adjustment of all three tuning currents. The control cir-cuitry required for the two laser types is almost identical andis described in Section VI-B. Newer versions of the SG-DBRlaser also include an additional amplifier section that boosts theoutput power above 10 dBm and an electroabsorption modulatorfor bit rates up to 2.5 Gb/s.

Another promising design is the digital supermode (DS) DBRlaser [26]. This type has multiple front DBR reflectors that canbe coarsely (digitally) adjusted to select one of the reflectionpeaks of a sampled grating that is located at the rear. The con-trols for this device differ from the previous two in that morecurrents have to be provided, but some of them do not requirethe same accuracy.

There are other designs that can meet the tuning speedand wavelength range requirements, such as array-wave-guide-grating-based multifrequency lasers (MFLs) [27], [28]and external cavity laser arrays [29]. Unfortunately, MFLs arecurrently not commercially available and external cavity laserarrays have drawbacks, such as large size. Integrated laserarrays that are currently under investigation [30] may offer acompetitive alternative to multisection DBR lasers in the future.

B. Control Electronics

Both the SG-DBR and the GCSR lasers are wavelength tunedby applying appropriate analog currents to the three tuning sec-tions. Due to manufacturing variations, these tuning currentsvary from one laser to the next. Typically, the manufacturerscalibrate each device and create a table with the tuning currentsettings for wavelength channels on the ITU grid. The man-ufacturers sell the lasers as butterfly packages or as moduleswith control electronics. Unfortunately, the commercially avail-able modules have so far not been optimized for high-speedswitching applications and typically deliver switching speedson the order of milliseconds. Therefore, we have designed fastcontrol electronics and mounted the butterfly-packaged lasersonto them [31].

Fig. 7 shows a schematic view of the electronics. A field-pro-grammable gate array (FPGA) holds a lookup table with threecurrent values for each wavelength channel. It interfaces to acontrol bus from which it obtains channel requests and a trigger

Fig. 8. (a) First-generation and (b) second-generation transmitter modules.The boards are shown at the same scale.

for switching. When the bus sends a trigger, the FPGA readsthe channel data and sends the corresponding current values to10-bit digital-to-analog converters (DACs). The DACs generatethree analog currents that are amplified and sent to the tuningsections of the lasers. The output impedance of the current am-plifiers is critical for fast switching, as we will discuss in Sec-tion VI-C and -D. The module also provides a gain current andtemperature control. The lookup table is based on the currenttables provided by the manufacturer and slightly modified in arecalibration procedure to match the conditions of the module.Also, since the manufacturers do not calibrate the devices forfast switching, we usually need to adjust the tables to obtain op-timum performance.

Fig. 8 shows two generations of laser modules. The first gen-eration (a) is a single printed circuit board (PCB), 18.5 cm10cm, on which the components are laid out similarly to Fig. 7.The second generation is significantly smaller (7.5 cm5 cm).This was achieved by stacking two double-sided PCBs on topof each other. The second module was designed to match thesize of commercial tunable laser modules and to be high-volumemanufacturable. There are multisource agreements for tunablelasers with slow tuning electronics [32], and this second gener-ation module could be a starting point for an extention towardfast tuning.

C. Impediments to Switching

The switching speed depends on the thermal and electricalproperties of the laser and the settling time of the control elec-tronics.

The currents applied to the tuning sections create local tem-perature variations due to resistive heating. Temperature tran-sients in these devices happen on the order of microsecondsto seconds and cause changes in the index of refraction in thetuning sections. This can lead to frequency shifts of the laseroutput of several gigahertz, depending on the design of the laser


Fig. 9. Setup for calibration and fast switching measurements.

and the magnitude of the tuning currents. Two common laserchip designs are the ridge waveguide and the buried heterostruc-ture. Generally the latter one has an advantage, since it requiressmaller tuning currents and removes heat more efficiently fromthe waveguide.

The electrical properties of the laser are determined bythe parasitic capacitance and resistance of the laser packageand the electrical properties of the laser chip itself. Standardbutterfly packages have fairly small parasitics compared to ourrequirements. Even though the manufacturers may not haveoptimized their package layouts for fast switching, the maincontribution stems from the chip itself. The tuning sectionsexhibit the electrical characteristics of forward-biased diodes.The dynamic impedance in each section is small as long as itis biased well above the electrical threshold but can be quitelarge close to threshold. As a result, lasers with larger tuningcurrents can be matched to high-impedance150 currentamplifiers without problems, whereas lasers with smallertuning currents and high capacitance show long settling times.To solve this problem, we modified the amplifiers of Fig. 7 intovoltage drivers to provide lower source impedance20 .This modification sacrifices the linearity between the digitalcurrent values and the actual tuning currents but allows for asignificant reduction of switching times even for lasers withsmaller tuning currents, as illustrated in Section VI-D.

The loss of linearity may at first seem like a penalty, but itactually carries another advantage beside increasing switchingspeed. The nonlinearity of theV–I curve of the tuning sectiondiode shows a trend similar to the nonlinear wavelength depen-dence on current. Thus, with the voltage driver the operationpoints of the laser are roughly equally spaced in DAC values,which provides efficient use of the 10-bit resolution.

D. Switching Measurements

We define the switching time as the interval betweenthe laser wavelength leaving channeland arriving at channel

to within both a specified frequency accuracyand a speci-fied power accuracy . The switching times between differentchannel combinations are typically going to be different andthe most important number is the maximum switching time

. For a laser that is calibrated for channels,there are 1 different switching times that need to bemeasured to determine .

Fig. 9 shows the setup used for laser calibration and fastswitching measurements. The light from the laser module is splitand sent to a wavemeter, an optical spectrum analyzer (OSA),and through a programmable tunable optical filter with 25-GHz

Fig. 10. Timing of the fast switching for an SG-DBR laser. Trace a) shows thedigital trigger signal as it enters the laser module. Trace b) shows the turnoff ofchannel 7 with a 12-ns latency. Trace c) shows the turnon of channel 23. It takes8 ns to switch to the new laser mode and 24 ns to reach the channel wavelengthwithin �8 0 GHz. Trace d) shows the optical power changing by 1 dB fromchannel 7 to channel 23. The power transient dies out before the wavelengthsettles.

FWHM into a photo receiver with 135-MHz bandwidth and a10–90% rise time of 2.2 ns. The photo receiver is connectedto a digital oscilloscope. A computer controls the laser moduleand the tunable filter and acquires data from all instruments.The wavemeter and the OSA are used for the calibration, thefilter, and the oscilloscope for the fast switching measurements.A frequency detuning of 8 GHz ( 12 GHz) corresponds to80% (60%) of the peak transmissivity for this particular filter.

Fig. 10 illustrates how an SG-DBR laser module switchesfrom one channel to another. The module receives an electronictrigger signal, shown as trace a), from the computer, and aftera deterministic delay of 12 ns, it switches from channel 7to channel 23. Trace b) was measured with the optical filtercentered on channel 7, and trace c) with the filter centeredon channel 23. The signals show that it takes about 24 ns forthe laser to make this particular transition. The laser switchesto the new laser mode in 8 ns, but due to current settling andthermal effects, it takes another 16 ns for the wavelength toreach the final value within the specified frequency uncertainty

GHz. Note that the laser power, shown as trace d),varies by 1 dB from channel 7 to channel 23, but the powertransient is short enough to have a negligible effect on theswitching time measurement.

Next we measure respective traces for allchannel combinations. We observe that the position of theturnoff [trace b)] is stable with respect to the trigger signal towithin a nanosecond, independent of what start and destinationchannels we choose. On the other hand, the turnon signalmoves with respect to the trigger depending on the particularswitching combination. We measure the time position of theturnon signals for all 992 combinations and plot the resultingswitching-time distribution in a histogram. The thresholds thatwe use for locating the falling and the rising edges correspondto the transmission percentages and determine the frequencyaccuracy of the measurement.


Fig. 11. Switching histogram for a) GCSR laser with high impedance currentdrivers, b) the same GCSR laser with low impedance voltage drivers, andc) SG-DBR laser with high impedance current drivers.

Fig. 11(a) shows the switching histogram for a GCSRlaser with small tuning currents and control electronics withhigh-impedance current drivers. Most switching times areunder 50 ns, but there are many combinations that take longer.The longest switching time is 80 ns for this configuration.Fig. 11(b) shows results for the same laser, but now witha low-impedance amplifier in the control electronics. Allswitching times are below 45 ns. This demonstrates the effec-tiveness of low-impedance amplifiers in reducing the settlingtimes of the tuning currents. The frequency accuracy for bothmeasurements was10 GHz.

Fig. 11(c) shows results for an SG-DBR laser. All switchingtimes are below 40 ns, even though the control electronicscontained high-impedance current amplifiers. This is consistentwith the fact that this particular laser requires higher tuning cur-rents that keep the bias on the tuning section diodes well abovethreshold and that it has lower chip capacitance. The frequencyaccuracy for this measurement was12 GHz. We had to relaxthe frequency accuracy in this case, because the higher tuningcurrents produced stronger thermal effects. This points to atradeoff in the design of these multisection DBR lasers. Highertuning currents facilitate obtaining high switching speeds, atthe expense of decreased switching accuracy.

The results of the switching measurements show that bothlaser types, the GCSR and the SG-DBR, are promising candi-dates for application in fast optical switch fabrics. The resultspresented in the rest of this paper focus on GCSR lasers, sincewe had only assembled and packaged modules with this lasertype at the time the experiments took place. Section VI-E ad-dresses the question of whether the laser modules are reliableenough for prolonged and continuous use.

E. Reliability Measurements

In Section VI-D, we established that GCSR and SG-DBRlasers can be used for fast switching between 32 channels in less

Fig. 12. Apparatus used for the reliability measurements.

than 50 ns. By measuring every switching combination once, wehave demonstrated the switching speed that these devices arecapable of, but we have not shown that these modules will actu-ally perform reliably in a realistic system environment. A tun-able transmitter module operating in a commercial switch fabricmay have to continuously switch once every microsecond witharbitrarily varying schedules for years without a single transi-tion exceeding the allowed maximum switching time. A 10-minmeasurement of 992 transitions clearly does not provide enoughstatistics to make a statement about its reliability when requiredto perform trillions of switching operations.

Therefore, we devised a measurement that allows us to op-erate the modules under more realistic conditions, to perform amuch larger number of measurements, and to study the effectof different schedules on the switching [33]. The results of themeasurements are then presented in the form of a packet-lossrate (PLR). We define a packet as lost if switching fails to occurwithin a specified time window or if the frequency or powerdeviations of the laser light exceed the specified uncertainties

and during the extent of the packet. The packet-lossrate is consequently the probability of packet loss due to laserswitching failures.

The measurement setup is shown in Fig. 12. A computersends a periodic schedule and a trigger signal to the tunablelaser module and sets the center frequency of an optical filterwith 25 GHz FWHM. The light from the tunable laser is sentthrough the optical filter and a variable optical attenuator intoa photoreceiver with 135-MHz bandwidth. The output of thephotoreceiver is added to a gating signal from a pattern gener-ator by means of a 6-dB coupler and sent into the data input of abit-error-rate tester (BERT). The pattern generator also providesthe overall timing in the form of a phase-adjustable trigger forthe computer’s digital I/O board and a 40-MHz clock signal forthe BERT. Note that there is no data modulation on the laserlight. We are not using the BERT in the conventional way ofdata transmission experiments but to oversample the 1s dataenvelopes at 40 MHz to count how often the laser fails to switchfast enough and if it drifts out of range during the packet.

We program a schedule in such a way that the tunable laseralternates between a test channelon every even time slot andthe other 31 channels on the odd time slots. The order in whichthe other channels appear can be arranged in many differentways, and we investigate two special types of schedules: linearand random. For example, a linear schedule for test channels20 cycles through the other channels and has the form

. It repeats after timeslots. In a random schedule , the other 31 channels arechosen randomly to form a long sequence with 20 000 values.


Fig. 13. Timing diagram for reliability measurements. (a) Signal from thephotoreceiver. (b) Gating signal. (c) Summation of the first two signals going tothe BERT; dotted line indicates decision threshold. (d) Clock signal for BERT.Arrow (1) BERT measures a one-bit independent of the switching jitter withinthe allowed time window. Arrow (2) BERT counts at least one error from thispoint on if the laser does not switch fast enough.

Before running a particular schedule, the computer sets theoptical filter to the wavelength of the test channel. It also ad-justs the variable optical attenuator in such a way that the max-imum optical power arriving at the photoreceiver is the same,regardless of which one of the 32 channels is chosen as the testchannel. The data threshold on the BERT is set to a fixed per-centage of the maximum received signal. As a result, the BERTmeasures a stream of one-bits as long as the laser wavelengthlies within a frequency window around the test channel, ora stream of zero-bits if tuning accuracy or output power is notachieved.

Fig. 13 shows the overall timing of the measurement. Trace(a) shows the signal at the output of the photoreceiver. When thelaser transmits at the test wavelength, the signal is high; other-wise, it is low. Trace (b) shows the gating signal, trace (c) thesum of the first two signals sent to the BERT, and trace (d) theclock signal for oversampling the BERT.

The high value of the gating signal at the position of arrow(1) causes the BERT to always count a one-bit. The gatingsignal ensures that the BERT will not mistakenly count errorsdue to varying switching times that are within the allowedtime window. The decision threshold indicated by the dottedline is set to 80% of the maximum filter transmission, whichcorresponds to a laser frequency uncertainty of8 GHz. TheBERT is programmed to expect a stream ofone-bits followed by a stream of zero-bits, sincethe gating signal adds two bits to the packet. If the laser failsto switch within a given time window, then the BERT willmeasure at least one zero-bit at position (2) and count at leastone error for that packet. Also, if the laser wavelength deviatesby more than 8 GHz during the extent of the data packet,the BERT will count at least one error. From the measuredBER, we can thus obtain an upper limit for the actual PLR fora particular test channel

PLR BER (4)

where is the size of the switching time window andtheoversampling rate given by the number of clock cycles per

Fig. 14. Switch diagram of a GCSR laser for a random schedule. Discretetraces for transitions from the 31 other channels to channel 20 are visible. Thehorizontal arrow indicates the range of positions of the first sampling pointwithin the packet, corresponding to arrow (2) in Fig. 13. The gray rectangleshows the time range of the gated sampling point, marked by arrow (1) in Fig. 13.

packet, 40 in this case. We can measure the PLR as a functionof the time window by adjusting the phase of the trigger thatthe pattern generator sends to the computer. This effectivelymoves trace a) left or right.

Fig. 14 shows a zoom-in of the turn-on of a test channel. Thefigure can be thought of as the equivalent of an eye diagramin that it shows a high-persistence oscilloscope picture of alltransitions from the other channels to test channel 20 for therandom schedule . Because of the analogy to an eye di-agram, we refer to this figure as a switch diagram. The switchdiagram shows discrete traces, which points to the repeatabilityof each individual transition from one of the other channels tothe test channel.

The gray rectangle shows the time range of the gated sam-pling point, marked by arrow (1) in Fig. 13. The horizontal arrowin Fig. 14 marks the range of positions of the first point withinthe packet that is sampled by the BERT, marked by arrow (2) inFig. 13.

A complete measurement requires the use ofscheduleswith the test channel set to each one of the laser channel. Theresult is an average PLR curve that measures packet loss for allwavelength channels

PLR PLR (5)

Fig. 15 shows the resulting individual PLR curves and theaverage PLR curve (thick line) for a GCSR laser. We usedlinear schedules and scanned the switching time window foreach test channel from 32.5 ns to a value where the experimentran error-free for more than one hour, with 48 ns being themaximum. This corresponds to a PLR of less than 610 .The bottommost point for each individual PLR curve thereforeindicates an upper bound obtained for an error-free measure-ment. The actual PLR values for the bottom points of manycurves are most likely much lower, given the steepness ofthe curves up to that level. Of the 32 schedules, seven wereerror-free even at 32.5 ns. Therefore, the figure only shows theother 25 curves.


Fig. 15. Packet-loss rate curves of a GCSR laser. The thin curves showPLRS for switching to individual channels; the thick curve shows the averagepacket-loss rate.

Fig. 16. Structure of the data packets. The gray areas indicate the trainingpattern before the packet payload.

The average PLR (thick curve) has been computed accordingto (5). It is dominated by the two or three rightmost PLR curves.All measurements combined demonstrate more than 32 hours oferror-free operation, or 58 billion tested packets with the laserswitching in less than 48 ns. This corresponds to an overall PLRof less than 1.7 10 .

The results of these reliability measurements demonstratethat multisection DBR lasers can be used in an environmentthat includes prolonged switching with different schedules. 32hours of error-free operation are certainly not the last word onlong-term laser performance, given that commercial systemshave to run continuously for months and reliably over years.Also, future work will have to include accelerated aging andthermal testing. The measurements do, however, present animportant first step in the direction of establishing guidelinesfor tunable lasers. This is all the more pressing, since existingTelcordia qualification procedures are solely designed for fixedwavelength lasers and new ways of qualifying tuning reliabilityhave to be established. We envision extended measurements ofthe kind presented here to play a role in future qualificationtests for tunable lasers.

VII. B URST-MODE RECEPTION

A third critical element of optical switch fabrics, besidesAWGs and fast tunable lasers, is burst-mode reception. It posestwo main challenges in the form of power and phase variationsfrom packet to packet.

Fig. 16 shows the basic structure of data packets as they ar-rive at the receiver. The packets are separated by dead times oflength . These are the times during which the lasers switchwavelengths. The packets start with training patterns of length

that allow the receiver to adjust its decision threshold to thenew optical power level and to recover the phase of the new databits. Then follows the data payload of length. The overhead

has to be kept small to make efficient use of theswitch fabric.

A. Power Variations

The receivers are subject to power variations betweenpackets caused by nonuniformities of the transmitter powersand by path-dependent transmission losses. In the extreme case,power variations will cause errors if they exceed the dynamicrange of the receiver, i.e., if the lowest power data packetsapproach eye closure and the highest power packets comeclose to the overload of the photoreceiver. Errors may alreadyoccur for much smaller power variations if the eye crossing andsymmetry change from packet to packet. Power variations canbe minimized either in the optical domain, through gain-com-pressed optical amplifiers or fast SOA-based adaptive gaincontrol (AGC), or in the electrical domain through limitingamplifiers or fast AGCs in the receiver. Another solution is fastthreshold adaptation. We use this approach in the followingdemonstration since it can be realized in a relatively straight-forward manner through AC coupling of the photoreceivers.

B. Phase Variations

The clock phase delay of the incoming data bits varies arbi-trarily from packet to packet. The reason lies in the fact thatit is practically impossible to match all fiber lengths and tophase-lock all transmitters to achieve stable phase relationshipsat high bit rates such as 40 Gb/s. Therefore the receivers have tobe able to quickly recover the phase of each new packet, prefer-ably within a few bits.

A suitable solution to this problem is clock data recovery(CDR) based on a digital Alexander-type phase detector and aphase-locked loop (PLL) with an internal voltage-controlledoscillator. This CDR can be integrated on a silicon-germanium(SiGe) chip together with a 1:16 demultiplexer from 40 to2.5 Gb/s [34]. The chip can be controlled through a fastproportional feedback path that recovers the phase and a slowintegrating feedback path that maintains a stable clock duringthe dead time .

VIII. O PTICAL SWITCH FABRIC DEMONSTRATION

This section describes a packet-switching demonstrationbased on an optical switch fabric containing the previouslydescribed components [35]. Due to hardware restrictions,only four ports of the fabric were populated with 40-Gb/stransmitters and receivers, but all components support at least a32-channel configuration, with a throughput of 1.2 Tb/s.

A. 4 4 Switch Fabric With 40-Gb/s Ports

Fig. 17 shows the setup of the switch fabric. Four tunabletransmitters connect to input ports of a noncyclical 4242AWG, the device described in Section V-B. Four output portsof the AWG connect to four burst-mode receivers. The trans-mitters each consist of a tunable laser module, a Mach–Zehnder


Fig. 17. A 4� 4 switch fabric setup. DFF: delay flip-flop; TLM: tunable laser module; OFE: optical front end.

modulator (MZM) and an erbium-doped fiber amplifier (OA)to overcome the total link loss of about 14 dB.

The laser modules contain tunable GCSR lasers that are cal-ibrated to switch between 32 ITU channels in under 50 ns, asdescribed in Section VI-D. They receive schedules and triggersignals from a computer that is timed by a pattern generator.

The MZMs are driven single-sidedly by four 40-Gb/s datasignals. The signals are generated as 16 data tributaries at2.5 Gb/s by the pattern generator, multiplexed to 40 Gb/s inone 16:1 multiplexer (Mux), duplicated by means of two delayflip-flops (DFFs), and amplified by modulator drivers. Twoof the four electrical signals are inverted, and through properadjustment of the modulator biases, we obtain four noninvertedoptical data signals.

Four optical front-ends (OFEs) with fast threshold adapta-tion are each connected to a 1:16 demultiplexer with fast CDR(see Section VII-B). Each demultiplexer then sends a recov-ered clock at 1.25 GHz and one 2.5 Gb/s tributary to one portof a four-channel BERT. The BERT is forced by a signal fromthe pattern generator to resynchronize at the beginning of eachpacket. After each successful synchronization to the packet pay-load, it counts transmission errors during the remainder of thepacket.

B. Operation

For a pseudorandom bit sequence (PRBS) payload with 2-1pattern length, the BERT requires about 6000 bits to synchro-nize. Since we sample the data with a 1.25-GHz clock, thistranslates into 4.8 s. We therefore program the pattern gen-erator to form 10-s-long 40-Gb/s data packets preceded by100 ns of dead time and a short training pattern. The timingof the schedule and the trigger are adjusted such that all fourlasers switch a few nanoseconds after the beginning of the deadtime. The schedule establishes connections between all inputand output ports in a simple round-robin pattern.

Trace (a) in Fig. 18 shows the data packets at one output portof the AWG. The three other output ports show similar vari-ations between packets, with a maximum power variation of

Fig. 18. (a) Received optical data packets. (b) Electrical signal from the BERTindicating synchronization (high) or loss of synchronization (low).

about 2 dB. Trace (b) in Fig. 18 shows how the BERT synchro-nizes to the signal. A logic high level indicates synchronization;a low level synchronization loss. The signal shows that it takesabout 6 s for the BERT to successfully synchronize on a datapacket. This instrument-specific limitation is the reason for ourchoice of a 10-s packet length.

C. Results

Fig. 19 shows oscilloscope traces of the four demultiplexed2.5-Gb/s signals as they enter the BERT. The traces show theend of the previous set of data packets (a), the 100-ns dead timeduring which the lasers switch, the start of the training pattern(b), and the locking of the CDR circuits to the bit phases ofthe new packets (c). Since the 40-Gb/s training pattern is of theform 00 110 011 the demultiplexed 2.5-Gb/s tributary is aconstant stream of ones or of zeroes. The transitions between(b) and (c) show that the CDR adjusts the phase of the PLL tolock to the bits of the new packet, a process that is completedafter 40 ns. Further PLL optimizations have shown that the locktime can be reduced to a few nanoseconds.


Fig. 19. Demultiplexed 2.5-Gb/s data packet as they arrive at the BERT. (a)End of a data packet. (b) Beginning of the training pattern of the next packet.(c) CDR phase-locks on training pattern.

Fig. 20. BER curves for (a) a back-to-back measurement without switchingand (b) the switch fabric in operation.

To ensure that the results are not affected by a particular com-bination in which we connect the transmitters and receivers tothe AWG, we mechanically connected each receiver to each oneof the four output ports of the AWG and performedBER measurements. Group (b) in Fig. 20 shows the results.All combinations achieve a receiver sensitivity of better than

3 dBm for a BER level of 10 . It is important to note thatthe received optical power is an average of the power levels ofthe individual data packets. The group of four BER traces (a) isobtained from a reference measurement in which the transmit-ters and receivers are connected without the AWG and the lasersare not switching. They show a receiver sensitivity of about

8 dBm for a BER level of 10 . The 5-dB penalty is causedby power variations between data packets, by the filtering ef-fect of the AWG, and by the fact that the fast threshold ad-justment does not provide individual threshold optimization foreach data packet. The sensitivity under switching is 6 dB belowthe receiver overload of3 dBm. This means that there is amplemargin for error-free operation of the switch fabric.

IX. CONCLUSION

In this paper, we have shown why optical switch fabrics mayplay an important role for next-generation multiterabit packetrouters. We have presented several optical router technologiesand compared their advantages and disadvantages with respect

to reliability, scalability, and feasibility. Then we described apromising multiterabit packet router architecture based on pas-sive arrayed waveguide gratings and fast wavelength switching.We explained the requirements associated with each key compo-nent and demonstrated that they are capable of meeting them. Fi-nally, we demonstrated packet switching with an optical switchfabric populated with four 40-Gb/s ports and capable of a totalthroughput of 1.2 Tb/s. The results show that this architecturewill enable next-generation terabit routers built around a reli-able and inexpensive optical switch fabric.

ACKNOWLEDGMENT

The authors would like to thank M. Berger, P. Mayer, S. Chan-drasekhar, M. Kauer, D. Stiliadis, D. M. Andrews, K. Sherman,F. Peragine, K. Dreyer, and R. Mclellan for enlightened discus-sions and their invaluable support.

REFERENCES

[1] K. Kar, T. Lakshman, D. Stiliadis, and L. Tassiulas, “Reduced com-plexity input buffered switches,” inProc. HOTI VIII, paper 1.3.

[2] R. Ryf, J. Kim, J. Hickey, A. Gnauck, D. Carr, F. Pardo, C. Bolle, R.Frahm, N. Basavanhally, C. Yoh, D. Ramsey, R. Boie, R. George, J.Kraus, C. Lichtenwalner, R. Papazian, J. Gates, H. Shea, A. Gasparyan,V. Muratov, and J. Griffith, “1296-port mems transparent optical cross-connect with 2.07 petabit/s switch capacity,” inProc. OFC’01, 2001,postdeadline paper PD28.

[3] A. Neukermans and R. Ramaswami, “Mems technology for optical net-working applications,”IEEE Commun. Mag., vol. 39, no. 1, 2001.

[4] J. E. Fouquet, “Compact optical cross-connect switch based on total in-ternal reflection in a fluid-containing planar lightwave circuit,” inProc.OFC’00, 2000, pp. 204–206.

[5] T. Goh, M. Yasu, K. Hattori, A. Himeno, M. Okuno, and Y. Ohmori,“Low loss and high extinction ratio strictly nonblocking 16� 16 ther-mooptic matrix switch on 6-in wafer using silica-based planar lightwavecircuit technology,”J. Lightwave Technol., vol. 19, no. 3, pp. 371–379,Mar. 2001.

[6] D. J. Bishop, C. R. Giles, and G. P. Austin, “The lucent lambdarouter:Mems technology of the future here today,”IEEE Commun. Mag., vol.40, no. 3, pp. 75–79, 2002.

[7] P. Granestrand, B. Lagerström, P. Svensson, H. Olofsson, J. Falk, and B.Stoltz, “Pigtailed tree-structured 8� 8LiNbO switch matrix with 112digital optical switches,”IEEE Photon. Technol. Lett., vol. 6, pp. 71–73,Jan. 1994.

[8] H. Okayama and M. Kawahara, “Prototype 32� 32 optical switch ma-trix,” Electron. Lett., vol. 30, no. 14, pp. 1128–1129, 1994.

[9] K. Hamamoto, T. Anan, K. Komatsu, M. Sugimoto, and I. Mito, “First8� 8 semiconductor optical matrix switches using GaAs/AlGaAselectro-optic guided-wave directional couplers,”Electron. Lett., vol.28, no. 5, pp. 441–443, 1992.

[10] J. Sasaki, H. Hatakeyama, T. Tamanuki, S. Kitamura, M. Yamaguchi,N. Kitamura, T. Shimoda, M. Kitamura, T. Kato, and M. Itoh, “Hybridintegrated 4� 4 optical matrix switch using self-aligned semiconductoroptical amplifier gate arrays and silica planar lightwave circuit,”Elec-tron. Lett., vol. 34, no. 10, pp. 986–987, 1998.

[11] N. Sahri, D. Prieto, S. Silvestre, D. Keller, F. Pommereau, M. Renaud, O.Rofidal, A. Dupas, F. Dorgeuille, and D. Chiaroni, “A highly integrated32-SOA gates optoelectronic module suitable for IP multi-terabit opticalpacket routers,” inProc. OFC’01, 2001, postdeadline paper PD32.

[12] D. Chiaroni, P. Bonno, O. Rofidal, J. Jacquinot, P. Poignant, C. Coeur-joly, F. Fernandez, E. Mestre, J. Moncelet, A. Noury, A. Jourdan, T.Zami, A. Dupas, M. Renaud, N. Sahri, D. Keller, S. Silvestre, G. Eilen-berger, S. Bunse, W. Lautenschlaeger, and F. Masetti, “First demonstra-tion of an asynchronous optical packet switching matrix prototype formultiterabit-class routers/switches,” inProc. ECOC’01, 2001, postdead-line paper ThA4.11.

[13] E. Shekel, A. Feingold, Z. Fradkin, A. Geron, J. Levy, G. Matmon, D.Majer, E. Rafaely, M. Rudman, G. Tidhar, J. Vecht, and S. Ruschin,“64� 64 fast optical switching module,” inProc. OFC’02, 2002, paperTuF3.


[14] M. K. Smit, “New focusing and dispersive planar component based onan optical phased array,”Electron. Lett., vol. 24, no. 7, pp. 385–386,Mar. 1988.

[15] C. Dragone, “AnN�N optical multiplexer using a planar arrangementof two star couplers,”IEEE Photon. Technol. Lett., vol. 3, pp. 812–815,Sept. 1991.

[16] P. Bernasconi, C. Doerr, C. Dragone, M. Cappuzzo, E. Laskowski, andA. Paunescu, “LargeN �N waveguide grating routers,”J. LightwaveTechnol., vol. 18, pp. 985–991, July 2000.

[17] K. Okamoto, T. Hasegawa, O. Ishida, A. Himeno, and Y. Ohmori,“32� 32 arrayed-waveguide grating multiplexer with uniform loss andcyclic frequency characteristics,”Electron. Lett., vol. 33, no. 22, pp.1865–1866, Oct. 1997.

[18] M. K. Smit and C. V. Dam, “PHASAR-based WDM-devices: Principles,design and applications,”IEEE J. Select. Topics Quantum Electron., vol.2, pp. 236–250, June 1996.

[19] P. G. Bernasconi, L. W. Stulz, M. A. Cappuzzo, E. Y. Chen, L. T. Gomez,E. J. Laskowski, and A. Wong-foy, “N�N arrayed waveguide gratingswith improved frequency accuracy,”IEEE J. Select. Topics QuantumElectron., vol. 8, pp. 1115–1121, Nov. 2002.

[20] J. Gripp, M. Duelk, J. Simsarian, S. Chandrasekhar, P. Bernasconi, A.Bhardwaj, Y. Su, K. Sherman, L. Buhl, E. Laskowski, M. Capuzzo,L. Stulz, M. Zirngibl, O. Laznicka, T. Link, R. Seitz, P. Mayer, andM. Berger, “Demonstration of a 1.2 Tb/s optical packet switch fabric(32� 40 Gb/s) based on 40 Gb/s burst-mode clock-data-recovery, fasttunable lasers, and a high-performance NxN AWG,” inProc. ECOC’01,2001, postdeadline paper ThA4.8.

[21] M. Duelk, J. Gripp, J. Simsarian, A. Bhardwaj, P. Bernasconi, M. Zirn-gibl, and O. Laznicka, “Fast packet routing in a 2.5 Tb/s optical switchfabric with 40 Gb/s duobinary signals at 0.8 b/s/Hz spectral efficiency,”in Proc. OFC’03, 2003, postdeadline paper PD08.

[22] F. P. Schäfer, W. Schmidt, and J. Volze, “Organic dye solution laser,”Appl. Phys. Lett., vol. 9, pp. 306–309, 1966.

[23] C. Shannon, D. Moore, and K. Claffy. (2002, Nov.) Beyondfolklore: Observations on fragmented traffic. IEEE/ACM Trans.Networking [Online]. Available: http://www.caida.org/outreach/pa-pers/2002/Frag/frag.pdf.

[24] V. Jayaraman, Z. M. Chuang, and L. A. Coldren, “Theory, design, andperformance of extended tuning range semiconductor lasers with sam-pled gratings,”IEEE J. Quantum Electron., vol. 29, pp. 1824–1834, June1993.

[25] R. C. Alferness, T. L. Koch, L. L. Buhl, F. Storz, F. Heismann, and M. J.R. Martyak, “Grating-assisted InGaAsP/InP vertical codirectional cou-pler filter,” Appl. Phys. Lett., vol. 55, no. 19, pp. 2011–2013, 1989.

[26] G. Busico, N. D. Whitbread, P. J. Williams, D. J. Robbins, A. J. Ward,and D. C. J. Reid, “A widely tunable digital supermode DBR laser withhigh SMSR,” inProc. ECOC’02, 2002, paper 3.3.2.

[27] J. B. D. Soole, K. Poguntke, A. Scherer, H. P. LeBlanc, C. Chang-Has-nain, J. R. Hayes, C. Caneau, R. Bhat, and M. A. Koza, “Multistripe arraygrating integrated cavity (MAGIC) laser: a new semiconductor laser forWDM applications,”Electron. Lett., vol. 28, no. 19, pp. 1805–1807,Sept. 1992.

[28] M. Zirngibl, C. H. Joyner, L. W. Stulz, U. Koren, M. D. Chien, M. G.Young, and B. I. Miller, “Digitally tunable laser based on the integra-tion of a waveguide grating multiplexer and an optical amplifier,”IEEEPhoton. Technol. Lett., vol. 6, pp. 516–518, Apr. 1994.

[29] M. Kauer, M. Girault, J. Leuthold, J. Honthaas, O. Pellegri, C. Goullan-court, and M. Zirngibl, “16-channel digitally tunable packet switchingtransmitter with sub-nanosecond switching time,” inProc. ECOC’02,2002, paper 3.3.3.

[30] Y. Kotaki and K. Morito, “Wavelength tunable DFB laser array forWDM applications,” inProc. ECOC’02, 2002, paper 3.3.1.

[31] J. E. Simsarian, A. Bhardwaj, J. Gripp, K. Sherman, Y. Su, C. Webb,L. Zhang, and M. Zirngibl, “Fast switching characteristics of a widelytunable laser transmitter,”IEEE Photon. Technol. Lett., vol. 15, pp.1038–1040, Aug. 2003.

[32] Multi-Source Agreement for Tunable Lasers. [Online]. Available:http://www.tunablemsa.com

[33] A. Bhardwaj, J. Gripp, J. E. Simsarian, and M. Zirngibl, “Demonstrationof stable wavelength switching on a fast tunable laser transmitter,”IEEEPhoton. Technol. Lett., vol. 15, pp. 1014–1016, July 2003.

[34] M. Reinhold, C. Dorschky, E. Rose, R. Pullela, P. Mayer, F. Kunz, Y.Baeyens, T. Link, and J.-P. Mattia, “A fully integrated 40-Gb/s clockand data recovery IC with 1:4 DEMUX in SiGe technology,”IEEE J.Solid-State Circuits, vol. 36, pp. 1937–1945, Dec. 2001.

[35] J. Gripp, M. Duelk, J. Simsarian, P. Bernasconi, A. Bhardwaj, K.Sherman, K. Dreyer, M. Zirngibl, and O. Laznicka, “4� 4 demonstra-tion of a 1.2 Tb/s (32� 40 Gb/s) optical switch fabric for multi-Tb/spacket routers,” inProc. ECOC’02, 2002, postdeadline paper PD2.4.

Jürgen Gripp (M’98), photograph and biography not available at the time ofpublication.

Marcus Duelk (M’96), photograph and biography not available at the time ofpublication.

John E. Simsarian, photograph and biography not available at the time of pub-lication.

Ashish Bhardwaj (M’02), photograph and biography not available at the timeof publication.

Pietro Bernasconi(M’96), photograph and biography not available at the timeof publication.

Oldrich Laznicka, photograph and biography not available at the time of pub-lication.

Martin Zirngibl, photograph and biography not available at the time of publi-cation.

Optical switch fabrics for ultra-high-capacity IP routers...

Documents

Transcript of Optical switch fabrics for ultra-high-capacity IP routers...