IEICE TRANS. ELECTRON., VOL.E82–C, NO.8 AUGUST 1999
1361
INVITED PAPER
Joint Special Issue on Recent Progress in Optoelectronics and Communications
Multiwavelength Opaque Optical-Crossconnect Networks∗ Evan L. GOLDSTEIN†a) , Lih Y. LIN† , and Robert W. TKACH† , Nonmembers
SUMMARY Over roughly the past decade, the lightwaveresearch community has converged upon a broad architectural vision of the emerging national-scale core network. The vision has been that of a transparent, reconfigurable, wavelength-routed network, in which signals propagate from source to destination through a sequence of intervening nodes without optoelectronic conversion. Broad benefits have been envisioned. Despite the spare elegance of this vision, it is steadily becoming clear that due to the performance, cost, management, and multivendorinteroperability obstacles attending transparency, the needs of civilian communications will not drive the core network to transparency on anything like a national scale. Instead, they will drive it to ‘opaque’ form, with critical reliance on optoelectronic conversion via transponders. Transponder-based network architectures in fact not only offer broad transmission and manageability benefits. They also make networking at the optical layer possible by offering to the nodes managed and performance-engineered standard-interface signals that can then be reconfigured for provisioning and restoration purposes by optical-layer elements. Because of this, the more pressing challenges in lightwave networking are steadily shifting towards the mechanisms that will be used for provisioning and restoration. Among these are mechanisms based on free-space micromachined optical crossconnects. We describe recent progress on these new devices and the architectures into which they fit, and summarize the reasons why they appear to be particularly well-matched to the task of provisioning and restoring opaque multiwavelength core long-haul networks. optical networks, optical crossconnects, optical key words: switching, WDM network architecture
1.
Introduction
WDM point-to-point transmission is fast emerging as the structural cornerstone upon which national-scale core communications will be built. Less clear, though, is whether the nodes of these WDM networks will be transparent—whether signals will propagate from source to destination through intervening nodes without optoelectronic conversion. Notwithstanding the past decade’s momentum in favor of transparent WDM networks, we argue that civilian communications needs will in fact drive the core network not to transparent, but to opaque form [1], with critical reliance on the emerging class of small, inexpensive optoelectronic converter known as the transponder. Manuscript received December 25, 1998. Manuscript revised February 15, 1999. † The authors are with AT&T Labs-Research, 100 Schulz Drive, Red Bank, N.J. 07701, U.S.A. a) E-mail:
[email protected] ∗ This paper is also published in the IEICE Trans. Commun., Vol. E82-B, No.8, pp.1095–1104, August 1999.
There are two broad reasons for this: (1) transparency is on balance a liability in national-scale networks; (2) the alternative to transparency (call it ‘opacity’) is on balance a virtue in such networks. These facts, once appreciated, lead one naturally to a broad class of opaque optical-crossconnect networks whose optical-transmission, network-management, and interoperability features make them, unlike transparent systems, eminently engineerable on a national scale. 2.
Opaque Networks
There are two long-standing arguments for building transparent WDM networks. The first concerns formatindependence: transparent WDM networks can transport analog or digital or unspecified modulation formats on separate wavelengths of a single network. This is clearly true in principle, and it constitutes an elegant feature of WDM optics. It is our working assumption, however, that the case for network-transparency will not rest on this form of format-independence. Global momentum towards digital intensity-modulated systems in general, and towards SONET and SDH protocols and bit rates in particular, has fast become overwhelming. What is critical instead is that a single network be capable of transporting a diversity of services— SONET, ATM, and IP-based—at a restricted set of specified standard bit rates. The second argument concerns node-bypass: transparent WDM networks readily allow express signals to bypass extensive electronic processing in intermediate nodes. This is clearly true for nodes consisting of a transparent fixed wavelength-add/drop-multiplexer, as shown in Fig. 1(a). It is a relatively small step to add reconfigurability, as in Fig. 1(b). From a transmission point of view, it is a substantial yet natural step to consider networks of transparent, reconfigurable wavelength-selective crossconnects, as in Fig. 11(c). All of the depicted network elements provide node-bypass, with its attendant prospects for cost-reduction. Transparency is thus an appealing feature: it allows signals to propagate at high power levels through long chains of losses. But transparency also lets small degradations do likewise. This opens the system up to a familiar, but ever-expanding, menagerie of cumulative performance-degradations due to chromatic and polarization-mode dispersion, optical-fiber
IEICE TRANS. ELECTRON., VOL.E82–C, NO.8 AUGUST 1999
1362
Fig. 1 Transparent implementations of a fixed wavelengthadd/drop-multiplexer (a); a reconfigurable wavelength-add/dropmultiplexer (b); and a reconfigurable wavelength-selective crossconnect (c). All offer the benefits of node-bypass.
nonlinearities, polarization-dependent loss, multipath interference, wavelength-misalignment of lasers and WDM filters, WDM-filter passband-narrowing, component crosstalk, noise-accumulation in ideal, flat-gain optical amplifiers, and a variety of gain-shape, crosssaturation, and compression effects in fiber amplifiers. These accumulating impairments have lately been amply chronicled by a voluminous literature (e.g., [2]– [6]), to which we add three observations. First, even when considered in isolation from one another, the above effects impose on the components of transparent networks stringent requirements that scale with reach. In-band crosstalk alone, for example, in a network corrupted by a single leakage path and operating at 2.5 Gb/s per wavelength, requires component crosstalk values smaller than −20 dB [4]. This is itself not a trivial demand, but it steadily tightens with system reach. Thus, a network corrupted by 32 leakage paths—not a large number in a national-scale system—requires component crosstalk values smaller than −40 dB (Fig. 2) [7]. For each of the other impairments listed above, a transparent network will suffer similar accumulation, and thus a similarly relentless tightening of component requirements. Transparency in WDM networks in general carries either performance costs or dollar costs or both, and these are offset only by savings in optoelectronic converters. This is evident from the accumulation of spontaneous-emission noise alone. Consider a transparent network whose amplifiers are ideal, with quantumlimited noise figures of 3 dB, flat gain spectra that precisely compensate fiber loss, perfectly equalized input-signal power-levels, and fairly aggressive amplifier output-power levels of +7 dBm/channel. Assume an operating margin of 6 dB. In such an idealized system, due to spontaneous noise-accumulation alone, the
Fig. 2 Scaling of in-band component-crosstalk requirements with number of leakage paths N . Experimental points, from Ref. [7], were obtained for signal and crosstalk having matched polarization and optical center frequency, at a per-wavelength bit rate of 2.5 Gb/s. Curves result from simplified Gaussian theory. Sources of degradation other than crosstalk are excluded.
Fig. 3 Decline of supportable per-channel capacity with total system length, due to amplifier noise, in a transparent network with ideal optical amplifiers (after Ref. [6]). The amplifiers are assumed to have ideal 3-dB noise figures, flat gain precisely compensating inter-amplifier loss, perfectly equalized signal-inputpower levels, and output power-levels of +7 dBm/channel.
supportable per-channel capacity declines with system length as shown in Fig. 3. Thus, a network with ideal amplifiers spaced at 80 km, operating at 10 Gb/s per channel, will suffer intolerable noise-build-up after about 3,200 km if the nodes are loss-free; finite node loss will necessitate proportionate reach-reduction. Further cumulative impairments suffered by real amplifiers, including transient cross-saturation, intolerance to lossvariations, and non-flat gain spectra, will tighten these constraints [6]. If one wishes to scale a transparent network beyond these limits, there are only two options. The first option is to accept degraded noise performance. The second option is to decrease the repeater spacing below 80 km, as indicated in Fig. 3. Network operators have thus far resisted this on cost grounds. Finally, the above comments na¨ıvely consider only
GOLDSTEIN et al: MULTIWAVELENGTH OPAQUE OPTICAL-CROSSCONNECT NETWORKS
1363
a single impairment acting alone. In a deployed national-scale transparent network, however, a small zoo of cumulative impairments will act in concert. Recent research shows that this raises two serious concerns. First, it has been found that two types of cumulative degradation that are individually benign can, when acting in concert, impose unexpectedly severe impairments. For example, the nonlinear-indexinduced broadening in a long-haul network-segment, when transparently passed to a conventional-fiber localexchange network, has been observed to produce severe dispersive error floors. These arise even in a system operating at less than 15% of its theoretical dispersion limit [8]. Moreover, such floors have been observed even for signals whose nonlinear broadening was wholly indiscernible using laboratory-grade optical spectrumanalyzers. This latter observation suggests that it will be difficult to erect signal-quality interface standards that guarantee performance in transparent long-haul networks, while at the same time being verifiable using field-deployable optical diagnostic equipment. If this is so, it is unclear how national-scale transparent networking would be made to work in practice. Transparency in fact constrains the upgradability of the network in two ways. First, as noted above, a transparent network has by definition lower capacity than a regenerated network. Second, a transparent network must standardize technology choices at the outset, and cannot exploit, for example, developments in broader amplifier bandwidths or narrower channel spacings as they become available. Opaque networks by contrast can employ the best technology available in each system that is added, with no barrier to interconnecting systems with different technologies. The various transmission systems entering a node need not use the same channel wavelengths, number of channels, or even modulation formats. This graceful adoption of new technology while continuing to use the old—a sort of ‘technology transparency’—represents one of the most powerful features of opaque optical networking. We expect the above considerations to have the following impact. Both fixed and reconfigurable varieties of transparent wavelength-add/drop-multiplexing (Figs. 1(a) and (b) are currently on the brink of deployment over limited (not national-scale) domains. Reconfigurable wavelength-selective crossconnect networks (Fig. 1(c)), however, even without wavelengthtranslation, are expected to prove engineerable on a national scale only by sacrificing either network cost or network performance or both. Moreover, if transparent wavelength-translation is needed in such networks, this will require adding a feature that has not yet been demonstrated in plausible form in the laboratory. It is an obvious but too-often-overlooked fact that the full variety of functional network elements in Fig. 1 is readily implemented in opaque form, by incorporat-
ing transponders at the interfaces of each transmission span. Opacity in this form exhibits exactly one bad feature: it requires numerous transponders. Notwithstanding their virtues, transponders are admittedly less than ideal in two respects. First, they are not free, although their precipitously dropping price trajectories, due partly to miniaturization, show no signs yet of leveling off. Second, transponders retime and reshape and are thus bit-rate-specific: increasing the bit rate requires transponder upgrade. In all other respects, opacity is no vice in optical-crossconnect networks. It is instead a positive virtue. Opaque crossconnects arrest cumulative transmission impairments through effectively complete signal-waveform clean-up. They facilitate performancemonitoring and fault-detection by monitoring, as desired, any of various SONET signal-quality indicators (e.g., loss-of-signal and loss-of-frame) and paritycheck bytes, and communicating the results to the optical-crossconnect system. Moreover, opaque optical crossconnects ease the task of fault-location since, by contrast with trans-parent architectures, their transponders offer a means of generating the alarminsertion (AIS) signals upon which downstream-alarmsuppression ordinarily relies [9]–[11]. Beyond this, transponders not only ease the task of topology selfidentification, by providing a way to insert identifiers; they also provide the function of wavelength-translation at no incremental cost [1]. Finally, and perhaps more important than all of the preceding considerations, opaque networks provide standard, nonproprietary interfaces. Since such open interfaces arise at the boundary between each WDM transport system and the optical crossconnects on which its endpoints terminate, the critical issue of multi-vendor inter-operability effectively vanishes. Thus, for a broad variety of reasons, of which physicallevel transmission impairments are only the most obvious, it grows increasingly clear that practical core-longhaul networks will be deployed in opaque form. 3.
Basic Restoration Architecture
Having resolved to build long-haul networks in opaque form, one still faces the far more complex challenge of choosing a restoration architecture. In multiwavelength networks, there are two fundamentally divergent ways of restoring failures, distinguished by the locations at which one places the restoration-switching elements. One approach is to respond to failures by rerouting individual wavelengths, or wavelength paths. This is achieved by placing switching elements on the “node side” of the network’s wavelength-multiplexers, where they operate only on single-wavelength paths [11], [12]. The various affected wavelengths in a failed fiber may then be rerouted over a variety of restoration routes. Alternatively, one can respond to failures by
IEICE TRANS. ELECTRON., VOL.E82–C, NO.8 AUGUST 1999
1364
Fig. 4 Wavelength-multiplex-section-level restoration vs. wavelength-path-level restoration. Fig. 5 Sharing of restoration facilities at the wavelengthmultiplex-section level.
rerouting full sections of wavelength-multiplexed signals, or wavelength-multiplex sections, as depicted in Fig. 4. In this section we examine the potential benefits of restoration at the wavelength-multiplex-section (WMS) level, and point out the operational difficulties that face this approach. We focus in particular on the cost of such an approach, compared with wavelengthpath-level (WP) restoration, when applied to nationalscale long-haul WDM net-works. WMS-level restoration in general offers two strengths, independent of the network’s geographical scale. First, traffic is restored in large bundles (full wavelength-multiplex sections), thus reducing both computational complexity and restoration switch-fabric size. Second, restoration is carried out “on the network side” of the wavelength-multiplexer, thus avoiding the expense of duplicated transmitter and receiver line cards. However, WMS-level restoration also faces three obstacles: (1) due to its coarse-grained nature, the approach necessarily uses restoration capacity less efficiently than do WP-level alternatives; (2) WMSlevel approaches create optically transparent (unregenerated) domains within which it will be difficult to deploy multi-vendor transmission equipment; (3) WMSlevel restoration paths, being both unregenerated and long in reach, will require the insertion of additional regeneration in order to satisfy transmission-engineering constraints. Thus, in general, restoring at the WPlevel economizes on transmission facilities, including fiber amplifiers, while restoring at the WMS-level economizes on the aggregate costs of transmitters and receivers as well as cross-connects. 3.1 Network Assumptions We examine a WDM mesh network whose topology is representative of the national-scale long-haul network, with approximately 450 switching offices (nodes) interconnected by 550 links. The links employ optical transport systems (OTS) supporting eight wavelengths, each
modulated at 2.5 Gb/s. Both the transmit and receive sides of the OTS are assumed to terminate on transponders. Transmission-engineering rules are assumed to permit 360-km transmission reach for systems built on 120-km amplifier spacings, and 560-km reach for systems built on 80-km repeater spacings. The network is assumed to support traffic projections for the year 2002. In the WMS-level approach, OTS failures are detected by transponders, and result in reconfiguration of WMS-level restoration cross-connects, which then provide an alternate path between end-points of the failed OTS. The restoration crossconnect at each office, as shown in Fig. 5, has a size of 2d × 2d, where d is the number of optical transport systems terminating at this office. The restoration crossconnect can connect any of the failed service fibers to restoration facilities. These crossconnects are configured to allow the restoration facilities to be shared among multiple optical transport systems. By contrast, we assume that WP-level restoration reroutes each affected connection between its endpoints employing what has come to be called end-to-end or source-based rerouting [13]. We consider only single OTS failures, and assume that any such failure must be restorable. 3.2 Numerical Results Under the above assumptions, WMS-level restoration is, as expected, found to be wasteful of transmission facilities. It on average requires 1.27 km of restoration fiber for each km of service fiber for the national-scale mesh network. WP-level restoration, by contrast, consumes 0.54 restoration km per service km. This finding is consistent with previous studies of WP-level restoration using end-to-end rerouting [11], [12]. The reason for this difference is granularity. WP-level restoration is able to utilize capacity more efficiently both because
GOLDSTEIN et al: MULTIWAVELENGTH OPAQUE OPTICAL-CROSSCONNECT NETWORKS
1365
it can extract restoration capacity from underused service fibers, and because it can employ bifurcated routing techniques [14]. In addition, WMS-level restoration consumes additional regenerators, used to satisfy transmissionengineering rules on the long WMS restoration paths. We assume that these regenerators can only be placed at the offices. For the national-scale network outlined above, an average of 1.1 regeneration points are required in systems built on 80-km repeater spacings, and 2.1 such points per restoration path in systems built on 120-km repeater spacings. These numbers can be reduced, however, by engineering the network so as to allow restoration paths corresponding to link-disjoint service OTSs to share regenerators. We employed a greedy algorithm to increase the sharing of regenerator locations, subject to transmission-engineering constraints [14]. The algorithm starts by assigning a regenerator to the location that can be shared by the largest number of paths. For those paths that have a regenerator assignment containing the selected location, all other assignments that do not include the selected node are discarded. The algorithm continues to select regenerator locations by assigning them to the location with largest sharing until, eventually, all paths satisfy transmission-engineering rules. Computationally, this turns on solving the Maximum Independent Set Problem [15], as outlined in Ref. [14] When the greedy algorithm is applied to WMSlevel restoration, one obtains a 31% reduction in regeneration points for systems employing a 120-km repeater spacing. By contrast, this declines to a 19% reduction for systems with 80-km spacings. The improvement is larger for 120-km spacings because one has more alternatives for assigning regeneration points to a given path, and thus more options for sharing. On the other hand, WP-level restoration requires transponders at the receive and transmit sides of each office traversed by the restoration path, and it is this approach that consumes very large numbers of transponders. By comparison, WMS-level restoration reduces the number of transponders on restoration paths by 89% at 80-km repeater spacings, and by 83% at 120km repeater spacings. Given the above results, together with equipment costs, the relative economic merits of WMS-level and WP-level end-to-end restoration are readily calculated. Equipment costs are normalized to the cost of an optical amplifier (=1). Link costs include only the costs of lighting, with optical amplifiers, a fiber that is presumed to be already available. We assume that the same basic technology is used for both the WP-level and WMS-level cross-connect, so that costs per port for the two approaches are identical. With these assumptions, the total network restoration equipment cost is plotted in Fig. 6 as a function of the unit transponder
Fig. 6 Normalized restoration equipment cost as a function of unit transponder cost.
cost per port. Total cost is normalized to that of the most costly system plotted, namely WP-level restoration with unit transponder and cross-connect port costs of 0.2 and 0.5, respectively. Because transponder costs are so rapidly dropping, only parametrized trend statements are possible. At transponder costs on the order of 0.2 to 0.4, roughly their value as of early 1999, aggregate system cost is seen to be dominated by the transponders. Thus, WMS-level restoration offers the shortterm promise of substantial cost advantages. However, should transponder unit costs drop by an order of magnitude, as miniaturization trends would appear to suggest, this advantage largely disappears. Thus, neglecting the operational liabilities of WMS-level restoration summarized above, this approach offers the prospect of significant short-term savings. However, particularly given the swiftly dropping trajectories of transponder size and cost, the manageability, interoperability, and upgradeability associated with wavelength-path-based approaches make their adoption appear inevitable. 4.
Free-Space Micromachined Optical Crossconnects
The immediate implication of the above is that the swiftly changing landscape of WDM networking, and the precipitous rise in aggregate per-fiber bit rates, are rapidly imposing the need for a new network element: a high-port count optical or optical-layer crossconnect. These network elements will likely be chiefly used for network restoration, to begin with. Substantial provisioning value will likely emerge thereafter. Optical switching technologies offer the potential advantages of bit-rate transparency, low power consumption, small volume, and low cost. Nevertheless, the requirements in port count (on the order of 1,000 ports in 3 to 5 years) and loss budget represent deep challenges that have not yet been met by any current photonic switching technology. Although conventional mechanical switches can
IEICE TRANS. ELECTRON., VOL.E82–C, NO.8 AUGUST 1999
1366
achieve high optical quality, they are large in size and mass, and are thus relatively slow in switching speed. On the other hand, guided-wave solid-state switches, though compact, generally have high loss and high crosstalk. The inherent disadvantages of these technologies thus appear to limit their expandability to the port counts mentioned above. By contrast, micromachined free-space opticalswitching technology holds particular appeal in this application because it combines the advantages of free-space interconnection—low loss and high optical quality—with those of monolithic integrated optics, namely, compactness. Various small-scale (2×2) micromachined switches [16], [17] utilizing sliding micromirrors have been demonstrated. In addition, collimating optics and rotating micro-mirrors have also been proposed as a means of achieving high-density optical switches [18]–[20]. For the application of restoration and provisioning in core-transport lightwave networks, free-space micromachined optical switches (FS-MOS) with free-rotating hinged micro-mirrors are particularly attractive. This is because such applications do not require frequent switching, but do require very high reliability even for switch mirrors that remain in one switching state for extended periods on the order of years. Furthermore, the sub-millisecond switching times exhibited by FS-MOS devices are well-matched to the needs of restoration and provisioning in core-transport lightwave-communications networks. In the remainder of this paper, we review the design and performance of FS-MOS devices, and their particularly strong match to the application of optical-network restoration and provisioning. 4.1 Device Design and Performance The working principle of a matrix FS-MOS is shown in Fig. 7. The micro-actuated free-rotating mirrors are monolithically integrated on a silicon chip by means of surface-micromachining techniques. The collimated light is switched to the desired output port by rotating a selected mirror. Figure 8 contains a schematic drawing of the actuated mirror. Micro-fabricated hinges [21] anchor the mirror on the Si substrate. The modified interleaved hinges constitute the hinge joints, through which the mirror is connected to the translation stage by push rods. The translation stage is actuated by arrays of scratch-drive actuators (SDA’s) [22]. FS-MOS chips utilizing mirrors of the abovedescribed structure have been fabricated using the MCNC MUMPs fabrication process [23]. The epitaxial layers consist of one Si3 N4 layer for insulation; three polysilicon layers for ground-plane purposes (poly-0) and for mechanical structures (poly-1 and poly-2); two phosphosilicate glass layers for use as sacrificial material; and one gold layer for use in fabricating electrical contacts and mirror coatings. The mirror and the trans-
Fig. 7 Schematic drawing of the matrix free-space micromachined optical switch (FS-MOS).
Fig. 8 Schematic drawing of the micro-actuated free-rotating switch mirror.
lation plate are built on the second polysilicon (poly1) layer, and anchored to the substrate through hinge and guiding rail structures built on the third polysilicon layer (poly-2). The SDA’s are L-shaped polysilicon plates formed on poly-2. The hinge joints consist of interleaved poly-1/poly-2 structures, and are connected to the pushrods built on poly-2. After epitaxial growth, the sacrificial material is selectively etched by immersing the device in hydrofluoric acid, thus releasing the mechanical structures from the Si substrate. Figure 9 shows a top-view photograph of an 8×8 switch. The switch fabric occupies a 1×1 cm2 chip. The free-rotating micro-mirror structures described above offer precisely the kinds of switching speeds that core-transport-network crossconnects demand. Device measurements have shown submillisecond switching times when the SDA’s are actuated with a ±100 V square wave at a frequency of 500 kHz. The maximum mirror-rotation angle is defined by the dimension of the ground electrode under the translation stage. Figure 10 shows the measured results for switching time. The switching time is 500 µs for rotating the mirror from the OFF position to the ON position. To this, one must add a 200-µs delay between the application of the switching voltage and the onset of measurable optical-switching action. For
GOLDSTEIN et al: MULTIWAVELENGTH OPAQUE OPTICAL-CROSSCONNECT NETWORKS
1367
Fig. 11 Polarization-dependent loss (PDL) of FS-MOS versus wavelength. The device’s PDL is below the noise floor of the measurement system.
Fig. 9
Top view photograph of an 8×8 FS-MOS.
Fig. 12 Schematic illustration of the bridging operation in core-transport networks. Fig. 10
Switching response of the FS-MOS.
the first demonstration, the mirror is rotated down by pulling the translation plate back with polysilicon springs, resulting in an ON→OFF switching time of 560 µs. In practical systems, the springs would likely be replaced by bi-directional SDA’s. The device’s switching curve shows an extinction ratio of more than 60 dB. By employing fiber collimators for input and output coupling, losses as low as 3.1 to 3.5 dB are achieved for the shortest and longest optical paths in an 8 × 8 switch, respectively. Unlike guided-wave approaches, these devices tend to exhibit very low crosstalk values (