Optical wavelength conversion will be a key function in photonic networks. Therefore the development of effective and practical all-optical wavelength converters is subject to considerable attention. Here, the application is outlined and the performance of wavelength conversion by semiconductor optical amplifiers reviewed.

The Supercomputer Supernet (SSN) is a high-performance, scalable optical interconnection network for supercomputers and workstation cluster based on asynchronous, wormhole- routing switches. The WDM optical backbone extends the geographic coverage range from interdepartmental to campus and even to metropolitan areas with dynamically reconfigurable direct or multi-hop connections. The network provides very high-speed integrated services, supporting connection oriented, guaranteed bandwidth traffic as well as datagram traffic. The first networking level of the two-level SSN architecture is electronic and consists of crossbar meshes locally interconnecting work-stations, supercomputers, peripheral devices and mass memory. At a higher networking level, an optical backbone network supporting multiple wavelength division multiplexed (WDM) channels allows communication between devices connected to distinct crossbar meshes. In this paper, we focus on the protocols of the WDM optical backbone network and address the issue of integration of the electronic wormhole- routing LAN with the optical backbone.

We describe an all-optical network testbed deployed in the Boston area, and research surrounding the allocation of optical resources -- frequencies and time slots -- within the network. The network was developed by a consortium of AT&T Bell Laboratories, Digital Equipment Corporation, and Massachusetts Institute of Technology under a grant from ARPA. The network is organized as a hierarchy consisting of local, metropolitan, and wide area nodes tea support optical broadcast and routing modes. Frequencies are shared and reused to enhance network scalability. Electronic access is provided through optical terminals that support multiple services having data rates between 10 Mbps/user and 10 Gbps/user. Of particular interest for this work is the 'B-service,' which simultaneously hops frequency and time slots on each optical terminal to allow frequency sharing within the AON. B-service provides 1.244 Gbps per optical terminal, with bandwidth for individual connections divided in increments as small as 10 Mbps. We have created interfaces between the AON and commercially available electronic circuit-switched and packet-switched networks. The packet switches provide FDDI (datacomm), T3 (telecomm), and ATM/SONET switching at backplane rates of over 3 Gbps. We show results on network applications that dynamically allocate optical bandwidth between electronic packet-switches based on the offered load presented by users. Bandwidth allocation granularity is proportional to B-Service slots (10-1244 Mbps), and switching times are on the order of one second. We have also studied the effects of wavelength changers upon the network capacity and blocking probabilities in wide area all-optical networks. Wavelength changers allow a change in the carrier frequency (within the network) without disturbing the data modulation. The study includes both a theoretical model of blocking probabilities based on network design parameters, and a computer simulation of blocking in networks with and without wavelength changers. Theory and simulation are in good agreement, and the results allow classification of those optical networks where wavelength changers provide benefit.

A multi-dual ring connected shuffle network is an optical multichannel multihop architecture proposed for wide area networks (WAN). With a simple fixed routing algorithm, this network architecture can achieve better performance than the Perfect ShuffleNet. In this paper, we propose a simple adaptive routing scheme which can achieve an even better performance. The adaptive routing scheme can quickly disperse packets away from congested portions of the network. Unlike centralized routing algorithms, the distributed routing algorithm uses only local state information and does not require a priori knowledge of the traffic patterns. Also, in contrast with some adaptive routing algorithms proposed for the Perfect ShuffleNet, the routing algorithm can distribute some of the traffic over less busy channels without increasing the length of the path. Moreover, in the case of network link failure, the adaptive routing scheme can direct the traffic around the 'trouble' area, which makes the network survivable. Since the whole idea of the scheme is to distribute traffic evenly among all the channels as much as possible, it can reduce the maximum traffic intensity on each channel, thus decreasing the size of pre-allocated buffer. All these characters make it very suitable for optical network architectures. The deloading factor is used in assessing the performance. Static simulations are performed under the worst condition and the most likely scenarios. The results support the preceding statements. The ideas presented here may be used for other optical WAN architectures, as long as each metropolitan area network is multiconnected in a ring topology.

The paper illustrates novel proposals for medium access control protocols in all-optical communication networks based on WDM multichannel ring architectures providing one slotted channel for each destination node. In such networks, nodes are equipped with one fixed- wavelength receiver and one tunable transmitter. Three access protocols based on local status information are described. A channel inspection capability is assumed to be available for the implementation of the access protocols. Global fairness control algorithms derived from those adopted in the Metaring metropolitan area network are also proposed. Access delays and throughputs are taken as performance indices for a simulation-based comparison of the proposed protocols, in the case of a 16-node multiring with either balanced or unbalanced traffic. Simulation results show that the throughput limitations and the fairness problems induced by the network architecture can be overcome with relatively simple access protocols.

Asynchronous transfer mode (ATM) is rapidly becoming the transport mechanism of choice for the information superhighway, because it promises the bandwidth and flexibility needed for many voice, video and data service offerings. Some industry experts project that the required sizes for ATM switching equipment in the public-switched environment will reach the Tbps range by the beginning of the next decade. This paper analyzes the problems associated with controlling the flow of packets within a broadband ATM switch of this size. The analysis is based on the requirements of the growable packet switch architecture. The paper proposes a novel solution to the problem of hunting paths within an ATM packet switch network. The resulting control scheme is unconventional in two ways. First, it uses an out-of-band control algorithm instead of the more common self-routing approach. In particular, we explore the benefits of using a parallel processor as an out-of-band controller for a growable packet switch distribution network. The processor permits additional levels of parallelism to be added to the out-of-band control function so that path hunts can be performed for all N of the input ports within a single cell interval. The proposed approach is also unconventional because it uses free-space digital optics to guide signals between successive stages of the controller. The paper describes the underlying motivations for implementing an optical out-of-band controller for an ATM switch, and it also describes the logic within a controller node that has been fabricated using a hybrid Si CMOS/GaAs SEED technology. The node uses optical detectors (in GaAs), amplifiers and digital control logic (in Si), and optical modulators (in GaAs). Free-space optical connections between successive device arrays can be provided using either bulk optical elements or micro-optics, but the optical interconnects must provide massive fanout capability. An architectural analysis studying the feasibility of applying free-space optics in this proposed ATM switch controller also is presented.

The architecture of a terabit free-space photonic backplane for parallel computing and communications is described. The photonic backplane interconnects typically 32 printed circuit boards or multichip modules (PCBs or MCMs) and has a bisection bandwidth in the terabit/sec range, with each PCB/MCM receiving a fraction of the peak bandwidth. The backplane consists of a large number of parallel reconfigurable optical channels spaced a few hundred microns apart. The parallel channels can be organized as a multi-channel ring where the channel access protocols are implemented by smart pixel arrays. Smart pixel arrays are integrated opto-electronic devices with optical I/O and with electronic processing capabilities. A smart pixel array which supports multiple reconfigurable broadcast channels and which transports terabits of optical data per second (Tb/s) and filters and extracts hundreds of gigabits of electrical data per second (Gb/s) for each PCB or MCM is proposed. Projections for free-space optical backplane capacities over the next decade are outlined. Within a decade, free-space optical technologies have the potential to support aggregate capacities of the order of 10s of Tb/s. It is also shown that optical technologies have the potential to result in significant reductions in volume and increases in performance when compared to conventional electrical architectures.

In the past few years, the demand for telecommunications services beyond voice telephony has skyrocketed. For the growth of these services to continue at this rate, cost effective means of transporting and switching large amounts of information must be found. Although fiber optic transmission has significantly reduced the cost of transmission, switching high bandwidth signals remains expensive. While all electronic switching systems are certainly possible for these high bandwidth systems, considerable effort has been expended to reduce the cost of fiber optic connections between frames or racks of equipment separated by several meters. As an example, one can envision fiber-optic data links connecting the line units that receive and transmit data from the outside world with an electronic switching fabric. Optical data links, ODLs, can perform the optical to electrical conversions. Several of these optical data links can be electrically connected with electronic switching chips on a printed circuit board. As the demand for bandwidth increases, several hundred to several thousand optical fibers might be incident on the switching fabric. Discrete optical data links and parallel data links with up to 32 fibers per data link remain an expensive solution to transporting this information due to their per-link cost, physical size, and power dissipation. Power dissipation on the switching chips is high because of the need for electronic drivers for the high speed electrical interconnections between the switching chips and the data links. By integrating the O/E conversions directly onto the switching chips, lower cost and higher density systems can be built. In this paper, we describe preliminary results of an experimental optoelectronic switching network based on this lower cost solution. The network is designed to be part of an asynchronous transfer mode (ATM) network based on the Growable Packet Architecture. The switching chip consists of GaAs/AlGaAs multiple quantum well modulators and detectors flip- chip bonded to silicon VLSI circuitry. The optical system images the inputs from a two dimensional fiber bundle onto the switching chip, provides optical fan-out of the signals from the fibers to the switching chip, and images the outputs from the chip onto the fiber bundle.

The fabrication of an interface card to extract free-space beams from an optical backplane and deliver them to either photodetectors on a printed circuit board or to fiber, is described. Experimental results are presented for polymer waveguides with 45 degree end reflectors and refractive polymer microlens arrays used in the fabrication of the card. Results for 1 by 4 and 4 by 4 prototypes which intercept 4 and 16 free-space beams, respectively, are also discussed. The 1 by 4 prototype extractor card coupled 4 free-space beams on 500 micrometer centers to direct-dispensed polymer waveguides with 45 degree end reflectors and finally to 100 micrometer core multimode fibers. Total throughput losses (free-space to fiber) for each of the four channels were 2.1, 2.4, 2.5, and 2.8 dB. In order to implement a fully connected 4 by 4 extractor card, laser-written waveguides were developed which have smaller cross-sections thereby allowing more channels and standard 62.5 micrometer core fiber to be used. The second prototype card intercepts 16 free-space beams in a 4 by 4 array (500 micrometer centers) and couples them into two sets of eight laser-written waveguides (50 by 50 micrometer cross-sections), which in turn are coupled to two multimode fiber ribbon cables with 62.5 micron cores on 250 micrometer centers. Total insertion losses from free-space to fiber ribbon are typically 4 - 5 dB. The refractive microlens arrays, which are used to focus the light onto the waveguide end-reflectors, are fabricated using an extremely simple direct- dispense technique which achieves excellent quality and uniformity. Experimental results are reported.

Kautz and de Bruijn networks are applied to model the free-space optica' interconnection of data arrays. For this purpose, the Kronecker sum (KS) and the Kronecker product (KP) of these networks is mapped into the 3-D physicaI space. The properties of the KS and KP networks are analysed and discussed. A switch and graph preserving transformation of l-D de Bruijn networks into their 2-D networks (and vice versa) is presented. The realization/refinement of the de Bruijn graphs by optical interconnections generates shuffle networks and thus the KP of 1st order of de Bruijn networks equals 2-D shuffle networks. The hardware requirement for the generation of permutations is analysed. Keywords: Data arrays, de Bruijn, Kautz, d-dimensional shuffle, Kronecker sum, Kronecker product, switch and graph preserving transformation

Surface-normal photonic switches are very attractive because they can switch and process a number of broadband optical signals simultaneously. Recent progress on surface-normal switches based on liquid crystals, MQW-modulators and surface-emitting lasers is reviewed in the context of system applications for optical interconnections and photonic switching networks.

Proposals to use large amounts of optics inside complex digital systems such as switches and computers impose new opportunities and new constraints on the optical receiver to be used. Since the receiver may exist on the chip with the VLSI, constraints on layout area and power consumption are severe. Since the system may be local in nature, the use of multiple-beam receivers may be possible. Finally, locality also implies that timing information may be available, permitting the use of synchronous sense-amplifier-based optical receivers. Issues associated with these points are discussed as they relate to several different examples of smart- pixel receivers. Results include high-impedance FET-SEED optical receivers with 3 mW power consumption and 1 Gb/s operation, hybrid-CMOS (0.8 micrometer) transimpedance receivers operating at roughly 200 fJ/bit at 1 Gb/s and below 10 fJ/bit at 311 Mb/s consuming 8 mW, and sense-amplifier circuits consuming about 1 mW at 330 Mb/s and 100 fJ/bit.

The design of a set of optics for a free-space digital optical backplane is presented. The optical system is designed to interconnect smart pixel chips using reflective modulators flip-chip bonded to CMOS circuitry. It consists of arrays of diffractive minilenses forming a 4f system around successive polarizing beamsplitter cubes. The minilenses between the optical power supply and the modulators are transmissive and on-axis, while those for the detectors are reflective and off-axis and emulate microprisms for beamsteering. The minilenses are augmented by microlens arrays for final focusing onto the clustered modulator windows. The optical module described has low loss, is highly compact and using f/4 diffractive optics supports 515 smart pixels/cm² for a chip size of 1 cm², giving a connection density of 2060/cm². Given flexibility in chip layout, this can be raised to a connection density of 4500/cm². Detailed analysis shows that the system meets performance targets set by considering present electronic backplanes, while conforming to the fabrication limitations of diffractive optics.

The advent of large-scale, free-space, opto-electronic interconnections, as demonstrated in recent system prototypes, requires new sampling methods to reveal diagnostic information. Several factors contribute to the difficulty of probing optical communications channels without disrupting their operation. High-speed electronic connections to the chip periphery are not available in sufficient number and would contribute an undesirable thermal load. Electronic and optical physical contact probes would obscure many of the optical channels that are relayed to a common surface of the chip in current systems. Optical sampling provides the better method although many standard techniques are either too time consuming or complex to implement. We describe a tool we developed that delivers diagnostic information on a large number of high-speed, optical data channels simultaneously and operates analogously to the conventional sampling electronic oscilloscope. The optical oscilloscope is constructed using CCD cameras and video capture boards that are controlled by a software application resident in a personal computer. Sampling is based on a stroboscopic method of using short pulsed laser probe beam synchronized to a data stream to illuminate optical modulators within the optoelectronic circuit. We have demonstrated and discuss the tool's capability of simultaneously monitoring arrays of broadband optoelectronic devices operating at speeds from several hundred Megabit/s to a few Gigabit/s.

Digital optical switches are very practical because they can provide flexibility for changing polarization, wavelength, temperature, and device dimensions without sacrificing crosstalk performance. For the first time, a novel compact digital optical switch in InP-based material with Gb/sec speed capability is proposed and demonstrated.

Arrays of fast smart pixels are the building blocks of high speed, high throughput photonic switching systems and spatial light modulators. Our work describes the operation of smart pixels with picosecond mode-locked laser pulses, and addresses some of the issues in the fast operation of smart pixels: switching energy, switching speed, dynamic range.

Photonic systems are already commonly deployed for the efficient transmission of huge amounts of data in both long haul and access networks. It is generally assumed that optical technologies will also enable the construction of ATM switches of very high capacity. While research continues into devices and techniques aimed specifically at switching applications, this paper presents an approach which is based on direct re-use and adaptation of technology elements from transport, in the construction of a very high capacity ATM switch. The architectural principles are, broadcast and select, and the grouping of ATM cells into larger bursts, transmitted at very high speed. Passive optical techniques are very effectively employed to replicate, and distribute the signals, while a selection of active photonics, in combination with electronics provide the switching, buffering, and control functions. The switch architecture provides a rich field for the application of various optical devices to increase throughput, and reduce cost. At the same time, advances in electronic technology evolution, such as faster logic and memory circuits, can go hand in hand with new photonic developments to reduce the size and power requirements of very high capacity switching systems.

An optoelectronic matrix capable of routing broadband digital and analog signals is presented. Metal-semiconductor-metal (MSM) photodetectors are used as bias-switched analog crosspoints to provide a signal format independent routing path through the switch. Results are presented for digital signals up to a bit rate of 2.5 Gbps with CATV signals switched simultaneously through the same switch fabric. This optoelectronic switch architecture provides a unique signal format insensitive method of implementing broadband switch fabrics required for gigabit communication networks.

The efficient exploitation of the coming 1 GB/sec optical interconnects will require radical re- design of the communication protocols and programming models used for distributed systems. The dominant programming models, such as message-passing, are complex enough to require extensive software involvement for data copying, matching received messages with posted buffers, and storing received data. This processing must use relatively slow accesses to non- cacheable memory structures. Consequently, the inter-processor latency, relative to the byte transmission time, is effectively increasing. Faster interconnects will exacerbate this problem. Lowest latency communication over 1 GB/sec LANs will require communication protocols and programming models that are implementable entirely in special-purpose hardware. True shared memory programming models allow this, but are unfeasible for workstations which also function as stand-alone machines, and are also extremely difficult to optimize over many processors. A proposed programming model for distributed systems, termed the 'shared array architecture,' combines the simple and deterministic communication of shared memory with the node independence and optimizability of message-passing. In the SAA model, each workstation/process in the user's parallel job maintains a 'shared array' of blocks of memory. These blocks map directly to transmission packets, and are directly read/write accessible by other processes in the job. A corresponding array of 'tag' values describe the status of the shared memory locations. Simple hardware mechanisms assure protection of non-shared data. Some similar type of mechanism, which closely ties the user interface to the communication software and hardware, will be necessary to exploit the capabilities of coming high-bandwidth transmission technology.

A compact optical crosspoint switch is being fabricated at the University of Colorado. It consists of two identical smart pixels arrays (SPAs) which face one another. The arrays communicate using 64 optical channels in both forward and reverse directions. Each channel uses a vertical-cavity surface-emitting laser (VCSEL) which is collimated using a microlens. The collimated beam is routed to the appropriate photodetector on the facing array using a Fourier computer generated hologram (CGH). Each VCSEL has a separate hologram associated with it, allowing space variant interconnection between the facing arrays. Each SPA has an 8 by 8 array of VCSELs adjacent to an 8 by 8 array of photodetectors and smart pixel circuitry. These are integrated with the microlenses and holograms and packaged to form a compact and stable system, occupying several cm³. At present each smart pixel has several logic gates, but an optically similar system with approximately 1000 gates per pixel is under development. This type of interconnection is suited to board-to-board applications, and potentially offers high speed, low cost, high density bi-directional connections. The fabrication of the switch incorporates several novel integration techniques.

We describe the parallel optical test of a photonic first-in-first-out (PFIFO) buffer memory based on flip-chip CMOS/SEED optoelectronic technology. The PFIFO detects pages of 4 by 8 binary bits, stores them in a 32 bit deep buffer memory, and transmits them in a 16 by 2 output modulator array. All 32 I/O and memory channels functioned, with an average input power of 40 microwatts and a minimum output contrast ratio of 2:1.

The rectangular crossbar network is commonly used in switching networks due to its many advantages, such as easy connection-path rebuilding, wide-sense nonblocking network, and no crossover in its interconnection lines. However, this network has the major disadvantage of non-zero differential loss. In this presentation the cyclic crossbar network is presented to reduce this differential loss to zero and still maintain the original advantages. In addition, using holographic optical switches to implement this new network with no interconnection lines is discussed in detail.

In order to provide cost effective solutions for different applications in optical interconnection, different approaches are needed. Cost and reliability are important considerations. The wavelength region between 900 nm and 1000 nm is very attractive since cheap Si-detectors can be used, the GaAs substrate is transparent and the fabricated Qw laserdiodes and LEDs have high performances and high reliability. Parallel optical interconnects require high performance (low threshold, high yield), densely packed laser arrays using fiber ribbon. For long distance communication, more emphasis is laid on the power budget and coupling efficiencies. If in parallel optical interconnect the very high modulation speed of laserdiodes is not really needed, InAlGaAs LED-arrays may be used, because of their stability, robustness and the possibility to integrate diffractive lenses on the backside of the component, which makes the component suitable for free-space optical interconnect. The better performances of microcavity LEDs will enhance this option even more.