The need for high bandwidth data transmission over distances of up to 10 kilometers and a large number of nodes has been fulfilled by the introduction of Enterprise Systems Connection Architecture (ESCON^TM) and Fiber Channel standards. ESCON is an IBM^TM serial interconnect standard for interconnecting IBM-compatible mainframes and their peripherals at 200 Megabud serial transmission rate. Fiber Channel is a serial interconnection standard for interconnecting supercomputers, mainframes, workstations, Personal Computers, storage devices, graphics terminals, scanners and printers at a data rate of 132 to 1065 Megabaud. This paper discusses the details of the GaAs/CMOS chipset for ESCON/Fiber Channel applications. It also describes the measurement setup for evaluating the jitter parameters, the jitter results of the transmitter and the jitter tolerance of the receiver.

We report the implementation and characterization of microstrip broadband pulse shaping filters for multigigabit optical communications. These filters were specially designed for application where the dominant noise is signal dependent, such as optically preamplified receivers. They are easily fabricated, insensitive to manufacturing tolerances, have reasonable physical size, and are easily integrated with the other components of the optical receiver.

A mixed signal Application-Specific Standard Product (ASSP) which integrates custom phase-locked loop (PLL) based clock recovery and clock synthesis functions with configurable logic cells provides a two-chip solution for 266 to 1062 Mbit/s Fiber Channel applications. The chipset performs parallel-to-serial and serial-to-parallel conversion, 8B/10B encoding/decoding functions, framing, and primitive signal/sequence generation and identification. The on-chip clock synthesis PLL generates the high-speed serial clock from a low-speed reference. The on-chip clock recovery PLL is capable of synchronizing directly to incoming digital signals, while simultaneously retiming and regenerating the data stream. The chips are designed using an advanced one-micron bipolar process with `Turbo' cells resulting in substantially reduced static power and output skews. Models for the PLL functions have been developed for use in logic simulation platforms. These models allow full chip-level and system-level simulation capabilities.

If this paper, we discuss techniques for measuring polarization mode dispersion (PMD) effects in experimental systems. Because PMD varies with time due to changes in temperature, pressure, and fiber position, measurements of PMD effects in a typical experimental system may not show worst case events or give a true indication of the distribution of these effects over the lifetime of a system. Here we show that by temperature cycling a spool of fiber using an oven, we can quickly determine the distribution of PMD effects, generating over 1000 independent PMD samples in a few hours. This permits the experimental study of events occurring with less than a 10^-3 probability and, in systems with only linear distortion, extrapolation of the results to much lower outage probability. Experimental results are presented for both external and direct (chirp) modulation at a 2.5 Gbps data rate over a fiber with polarization dispersion with an average delay of 120 psec for direct detection.

Coaxial cable and distributed switches provide a way to configure high-speed Fiber Channel fabrics. This type of fabric provides a cost-effective alternative to a fabric of optical fibers and centralized cross-point switches. The fabric topology is a simple tree. Products using parallel busses require a significant change to migrate to a serial bus. Coaxial cables and distributed switches require a smaller technology shift for these device manufacturers. Each distributed switch permits both medium type and speed changes. The fabric can grow and bridge to optical fibers as the needs expand. A distributed fabric permits earlier entry into high-speed serial operations. For very low-cost fabrics, a distributed switch may permit a link configured as a loop. The loop eliminates half of the ports when compared to a switched point-to-point fabric. A fabric of distributed switches can interface to a cross-point switch fabric. The expected sequence of migration is: closed loops, small closed fabrics, and, finally, bridges, to connect optical cross-point switch fabrics. This paper presents the concept of distributed fabrics, including address assignment, frame routing, and general operation.

STARMAP is a new, active star-configured, multiple access protocol designed particularly for very high-speed fiber optic LANs but equally applicable to lower speed copper based systems. The main features include: collision-free operation; no packet retransmissions; bounded access delay time; high degree of service fairness; no back-off algorithm required; an integrated data/voice transmission capability; a Universal, a Selective and a Local (Global & Selective) Broadcast capability; very high security; Local Selective Broadcast packets never leave the local hub; a relative insensitivity to `Master' hub failure; preemptive and nonpreemptive priority packet service scheme; novel variable delay register in the hubs; excellent natural diagnostic capability; Loop Creating Links significantly improve network performance; true parallel transmissions. Computer simulations of example STARMAP networks show that at typical values of the offered traffic load, the network throughput exceeds the link bit rate and in the limit, approaches a value equal to the product of the link bit rate and the number of hubs in the network. The useful life of twisted wire pair and coaxial cable based networks may be significantly extended due to the substantial increases in network throughput achievable.

The Fiber Channel (FC) being defined by the ANSI X3T9.3 task group provides a general transport vehicle for a number of upper level protocols. Interconnection of Fiber Channel nodes is accomplished using the Fiber Channel Fabric which is being defined by the FC Fabric Special Subject Working Group (SSWG). This paper discusses the general concepts of a fabric, describes the various fabric topologies being considered, and discusses some of the characteristics of a fabric.

Current and emerging high-speed applications and environments create new and challenging requirements for high-speed channels on large systems. The Fiber Channel Standard (FCS) defines transmission and signalling protocols that allow various Upper Level Protocols (for example, IPI3 and IP) to be transported over a variety of physical interfaces. This rich set of function allow FCS to address many aspects and varieties of high-speed interconnect, by permitting a variety of implementations, based on trade-offs of cost, performance objectives, implementation complexity, and recoverability. The current presentation describes several high-speed applications and environments and the usage of FCS channels by large systems in those environments. Application of various FCS functions and features to this large system environment are discussed.

Enterprise Systems Connection Architecture ^TM (ESCON) is a new IBM^TM input/output (I/O) and interconnection architecture. Fiber Channel is an ANSI X3T9.3 draft standard which describes a high-performance serial link for support of higher-layer protocols associated with HIPPI, IPI, SCSI, and others. The two interfaces share many characteristics. This presentation focuses on the differences.

There exists an increasing need, in the user environment, for a computer interconnect scheme with higher speed, higher performance and longer reach than the presently available alternatives. There is also a great demand for a multidirectional networking to provide high bandwidth on demand, high distribution capability, random access and high transport flexibility. The users expect low access delay, low transfer delay, high data integrity and a definable quality of service from their networks. All these requirements, however, have to be met with the preservation of the existing software in which a lot of user investment has already been made. In answer to the demands, there has been an emergence of a new network to interconnect heterogeneous systems at very high cost performance ratio. This new network is based on Fiber Channel Standard, blessed by the American National Standards Institute (ANSI).

The University of Hawaii Institute for Astronomy is developing a high-bandwidth data communications system that will connect the telescope facilities of the Mauna Kea Observatories with base support facilities at lower elevation and with other institutions worldwide. A key component of this project is an extensive fiber-optic cable plant that links the observatories at the Mauna Kea summit with each other and with a mid-level support facility. The first application of the fiber-optic system, a fiber distributed data interface (FDDI) token ring with a circumference of over 32 km and seven nodes, is in operation. Plans are underway to install an OC-12 or OC-24 Sonet ring to improve the efficiency of fiber use. We describe the needs and applications of the multinational Mauna Kea Observatories, the current network configuration, impending network development, and future networking plans to accommodate additional users and applications. We summarize our experiences in dealing with FDDI token rings over single-mode fibers.

The IBM Enhanced Clustered Fortran (ECF) advanced technology project combines parallel computing technology with a HiPPI-based LAN network. The ECF environment is a clustered, parallel computing environment which consists of IBM ES/9000¹ complexes and possibly other parallel machines connected by HiPPI. The ECF software, including the language processor, is independent of hardware architectures, operating systems, and the Fortran compiler and runtime library. The ECF software is highly portable because it is based on well-known, standard technology and transport protocols such as Remote Procedure Call (RPC), X/Open Transport Interface (XTI), and TCP/IP. The ECF software is transport-independent, and can accommodate other transport protocols concurrently. This paper describes the IBM ECF environment including the language extensions, the programming model, and the software layers and components. Also, this paper explains how to achieve portability and scalability. Lastly, this paper describes how effective task communication is accomplished in ECF through RPC, XTI, TCP/IP, and a customized enhancement over HiPPI. An analysis of network performance in terms of bottleneck conditions is presented, and empirical data indicating improved throughput is provided. Comparisons to alternative methodologies and technologies are also presented.

The FDDI ring-protocol combination is well-suited for large local (`campus') area networks. However, for larger networks, FDDI users suffer excessive access delay, even under very light traffic. Moreover, FDDI achieves full capacity only in special circumstances, and heavy load is not a sufficient condition for achieving it. We propose a modification of FDDI that has none of these deficiencies. By changing the interpretation of holding the token, we get a protocol that allows every station to transmit when the link is silent. The resulting token-passing protocol yields a capacity-1 network even for modest values of TTRT. Additionally, stations endure a very small access delay under light load. All the other properties are identical to those of FDDI. The proposed protocol uses standard FDDI components. Moreover, the protocol is compatible with FDDI to the point that it works correctly even if some of the stations attached to it operate under the original FDDI protocol (those stations will get the performance offered by FDDI, while the stations using the new protocol will get a larger bandwidth and a lesser access delay).

Multichannel local-area networks can be constructed using fiber-optic wavelength-division-multiplexing (WDM) techniques. In this paper, WDM is applied to ring networks such as FDDI, token ring, slotted ring, and buffer insertion ring architectures. Crosstalk and insertion loss analyses of grating demultiplexers shows that the ratio of channel separation to channel width must be greater than four. Performance analyses of the multichannel WDM ring are formulated to determine the packet delay. The delay in the multichannel token ring is obtained by summing the remaining service time, token walk time, and the average service time of newly arrived messages. The remaining service time is estimated by approximating the system as a M/M/m queue. The multichannel slotted ring and buffer insertion ring are modeled by a non-preemptive head-of-the-line priority queuing system. By identifying that portion of the passing ring traffic having priority over local traffic, the delay of a tagged packet can be obtained as the sum of the residual service time and service times of local waiting packets and newly arrived ring packets. Two implementations are identified, one with parallel optical transmitters and receivers and protocol modules, and another using an array of laser diodes and photodetectors and a single protocol module.

This paper briefly describes the spectral sliced testbed. The architecture pre-allocates channels for data reception and results in a system with M nodes and C subsystem at each node, since only a single wavelength channel is delivered to each node. In addition to the low complexity, the proposed spectral sliced approach supports multicast and broadcast modes in which one processor can simultaneously transfer data to more than one destination. Protocol complexity is the major constraint in this environment. The protocols must have both low implementational and operational complexity to maintain practical feasibility at the high operating speeds. Pre-allocation approaches appear to be very promising because of their low implementation and operational complexity. Two pre-allocation protocols are introduced that are targeted to the spectral sliced architecture. The resulting performance of the two protocols is examined in terms of average delay and throughput through discrete event simulation.

Multi-domain wavelength-division multiplexing (M-WDM) is proposed as a generalization of single star-coupled WDM networks. The objective is to provide scalable interconnection schemes for different parallel computer architectures that utilize the multi-channel and tunability characteristics of photonic network devices, given their existing limitations. Device tunability results in reconfigurable and partitionable interconnection. The multi-domain structure results in scalable networks due to the spatial re-use of the available wavelength band. This paper identifies the variables involved in the design of (M-WDM) networks according to arbitrary interconnection schemes. A framework for network specification and parameter trade-off is presented. The introduced network structures are potentially applied to both synchronous and asynchronous parallel computing environments. Topologically reconfigurable architectures are presented for both single-stage and multi-stage (pipelined) parallel machines executing synchronous algorithms. M-WDM structures are compared for distributed shared memory, multiple-instruction multiple-data (MIMD), parallel architectures.

Wavelength Division Multiplexing (WDM) enables partitioning the enormous bandwidth of photonic networks into multiple smaller, more manageable, multiple access channels. These channels operate at a data rate which matches the electronic interface speed, viz. Gbps. Media access protocols for an optically interconnected star-coupled WDM network with no control channel are introduced and compared. The channels are preallocated to nodes where each node has a home channel that it uses for all data reception. If the number of nodes exceeds the number of channels, home channels are shared among nodes. This approach does not require both tunable transmitters and tunable receivers reducing system complexity and is not limited by the number of channels available. A generalized random access protocol and an interleaved time division multiplexed protocol are compared. Both protocols require a fast tunable transmitter and a slow (or fixed) tunable receiver per node. Each node has a set of queues of variable capacity -- one per data channel. The switching time of tunable transmitters has a significant impact on system performance and techniques are developed to reduce the impact. Detailed discrete-event simulation results are used to evaluate system performance in terms of network throughput and average packet delay with variation in the number of nodes and channels and transmitter switching latency.

This paper investigates the impact of fast wavelength tunable device switching latency on the performance of media access protocols for star-coupled wavelength-division multiplexed photonic networks. The relative impact with a reservation based protocol (TDMA-C) and two pre-allocation based protocols (I-TDMA^* and I-SA) are compared. TDMA-C is control channel based, with one WDM channel allocated to reserve access for data packet transmission on the remaining data channels. Control channel access arbitration is achieved through time-division multiplexing, enabling all active nodes the opportunity to transmit once very control cycle. I-SA and I-TDMA^* are designed for a network where channels are pre-allocated to the nodes for reception where each node has a home channel it uses for all data packet receptions. The performance of the protocols is evaluated through discrete-event simulation in terms of network throughput and average packet delay. In particular, this paper examines the performance impact with variations in the number of nodes and data channels, packet generation rate, data packet length, and optical device switching latencies.

This paper introduces a preallocation-based multiple access protocol for wavelength-division multiplexed (WDM) optically interconnected passive star-coupled networks. The objective is to combine the strengths of two other preallocation-based strategies, interleaved slotted ALOHA (I-SA) and interleaved time division multiple access with multiple buffers (I-TDMA^*), while minimizing their inherent weaknesses. I-SA is a random access protocol which features good delay but poor maximum throughput. I-TDMA^* is a static protocol which has excellent throughput but suffers from synchronization delay at low arrival rates. The new protocol, dynamic interleaved slotted ALOHA (DISA), is a random access protocol which combines the high capacity of I-TDMA^* with the excellent latency characteristic of I-SA. This is achieved through a flow control mechanism which throttles the amount of traffic transmitted through the network based on the local load characteristics. No global information (other than packet acknowledgments) is necessary to implement this flow control mechanism; each node intelligently restricts network transmissions based only on the frequency of local packet arrivals and collisions. The protocol is described and discrete event simulation is used to compare the performance of DISA to I-SA and I-TDMA^* over a range of system, load, and protocol parameters.

A single-hop wavelength division multiplexing (WDM) based local lightwave network employing a passive star topology is considered. The system consists of a single control channel and a number of data channels. Each station is equipped with a transmitter and a receiver, both of which are tunable over all the channels. Nodes employ the control channel to arbitrate (coordinate) their access to the data channels. The attractiveness of this architecture is its extreme simplicity. We have previously proposed a Receiver Collision Avoidance (RCA) protocol for such a system in which all nodes are equidistant from the passive star. Under the RCA protocol, access to the control channel is provided via a variation of slotted ALOHA, which includes a simple mechanism that can dynamically detect and avoid receiver collisions. The protocol is scalable and can support a large number of bursty nodes with a relatively small number of data channels. In this paper, we consider an extended RCA protocol (E-RCA) to incorporate nonuniform distances, while maintaining all of the protocol's original attractive features. The analytical model for the E-RCA protocol is difficult to formulate; therefore, extensive simulations were conducted under various distance distributions. Results indicate that the E-RCA protocol performs almost as well as the RCA protocol under the same average distance conditions. Also, like the RCA protocol, the E-RCA protocol is simple, based on practical assumptions, and can be readily implemented with current lightwave technology.

We present a time slot sorter which can sort N destination time slot numbers in N time slots of a frame in a system which uses time division multiplexing. The time slot sorter, S_N, is obtained from the spatial sorting network, the Batcher network, and consists of a series of 1/2n(n ^{+ 1} exchange switches with associated delay elements, N equals 2ⁿ. The individual switch settings of S_N are determined on-the-fly by pairwise comparison of destination time slot numbers included in the time slots entering each switch. A time slot sorter is more powerful than a time slot permuter in that it can sort N arbitrary destination time slot numbers whereas a time slot permuter can sort a permutation of N integers 0 through N -1. A time slot sorter has an important application in a general switching problem known as the time-space-time domain permutation in telecommunications. Obviously a time slot sorter can also be used as a time slot permuter.

We consider the general switching problem known as time-space-time domain permutations in telecommunications. We present a new set of multichannel time slot permuters for L parallel frames of M time slots (L equals 2^l, M equals 2^m). The multichannel time slot permuters are obtained by combining L X L spatial networks and time slot permuters for a frame of M time slots. In this paper, the Benes network, the Batcher sorter and the Lambda network for spatial networks, and their counterparts, the RJS time slot permuter, the S time slot sorter, and the Lambda time slot permuter are considered.

We present a new self-routing time slot permuter, the generalized Lambda time slot permuter (Lambda) _N^q for a frame of N time slots (N equals 2ⁿ, n equals pq), which generalizes the previously proposed Lambda time slot permuter. The generalized Lambda time slot permuter is obtained by combining the idea of the Lambda time slot permuter and Q-way bitonic decomposition (Q equals 2^q). Each instance of a switch setting in (Lambda) _N^q is determined on-the-fly by a q-bit comparison of input destination tags. The local control complexity of a generalized Lambda time slot permuter (Lambda) _N^q varies from n-bit comparison (q equals n) to a single bit control of the Lambda time slot permuter (q equals 1), while the hardware complexity varies from O(n² to O(n³). The complexity of the q-bit comparison is O(1) if the data is presented serially, but allows some trade-offs to be made if data packets and destination tags have some parallel nature.

We discuss the logical aspects of multigrid network architecture (MNA) -- a networking concept aimed at applications calling for very high transmission rates (of the order of gigabits per second). We also present a simple performance model applicable to regular topologies of MNA and demonstrate how to apply this model to sample network configurations. The technology for building MNA switches has been developed in Lockheed Palo Alto Research Lab. A MNA network is a collection of photonic switches linked by optical channels. A switch is responsible for relaying incoming packets along the best available routes to their destinations. Packets that cannot be relayed optimally are deflected, i.e., relayed along suboptimal paths.

One of the principal limitations in parallel real-time processing systems is the technique used to interconnect processors. Recent research has developed a novel optical transmission architecture to improve processor-to-processor connectivity bandwidth and permit reconfigurability of processor communications. Spectral sliced technology is used in this interconnect architecture to provide processor connectivity using a wavelength division multiple access addressing protocol. Spectral sliced technology employs the broad spectral output characteristics of light emitting diode (LED) sources and the spectral filtering characteristics of multiple channel WDM devices to provide a number of simultaneous non-blocking processor-to-processor connectivity paths. The spectral sliced interconnect technique can also support multicast and broadcast modes which permit one processor to simultaneously transfer data to more than one destination. This paper describes the architecture and provides measured results from an eight node prototype testbed configuration. A parametric study is presented that demonstrates the effects of source wavelength and spectral filter bandwidth on network performance and extent. These studies indicate that networks with high effective throughput rates can be constructed from commercially available off-the-shelf components.

Message congestion in heavily loaded mesh networks can limit system throughput. While electronic express routers increase mesh system throughput by alleviating congestion and reducing average message delivery time, optical express routers offer the possibility of a further order of magnitude performance increase by exploiting the latency advantages of decoupling data rate from optical router reconfiguration time. We show how a self-routing optical mesh can be interfaced to an existing electrical mesh, and describe the use of optimized default switch settings to minimize latency. We show that circuit switching exploits the benefits of optics in express routers to a greater extent than does packet switching, and illustrate likely architectures in which waveguide switches, detectors, and amplifiers are integrated to form self-routing nodes. The self-routing mesh algorithm and data format required to guarantee that the correct path is established are described. Our calculations show that for meshes of 512 nodes or more, optical express routers allow an order of magnitude increase in throughput compared to electronic express routers, and possibly two orders of magnitude compared to non-express meshes.

The throughput of virtual connections is negotiable in future broadband integrated services digital networks (B-ISDN). A new problem for admission control in broadband networks, thus, is to decide which calls to accept as well as at which throughput levels. This paper presents an optimization model and its solutions for the admission control and throughput negotiations in B-ISDN. The model is an extension of the knapsack problem, namely the flexible knapsack problem. It differs from the conventional knapsack problems in allowing different packing forms for the objects from the same class. Each packing form is associated with a different set of volume requirements and reward rates. The objective is to maximize the total reward. The decision space in the flexible knapsack problem is not only acceptance and denial, but also includes the choice of the optimal packing form for each accepted object. Three algorithms are provided in this paper to obtain the optimal access control decision for the connection throughput negotiations in B-ISDN.

High-speed networks and high-performance workstations are necessary but not sufficient to support distributed multimedia applications. A real-time scheduling system designed for multimedia data types is also required to orchestrate communications channels, disk storage units, output devices, and the CPU. These subsystems are coordinated to accommodate the special requirements of multimedia data: timely retrieval, transmission, and delivery with permissible levels of data loss and corruption. In this paper we present our framework for the use of real-time scheduling disciplines to support time-dependent multimedia data in a distributed-data environment. Within this framework we propose the application of a statistical resource reservation mechanism and a real-time session scheduler. Furthermore, we relate scheduling and quality of service in a summary of the objectives of multimedia service provision and negotiation.

We propose a modification to the Fiber Distributed Data Interface (FDDI) protocol based on a simple algorithm which will improve confidential communication capability. This proposed modification provides a simple and reliable system which exploits some of the inherent security properties in a fiber optic ring network. This method differs from conventional methods in that end to end encryption can be facilitated at the media access control sublayer of the data link layer in the OSI network model. Our method is based on a variation of the bit stream cipher method. The transmitting station takes the intended confidential message and uses a simple modulo two addition operation against an initialization vector. The encrypted message is virtually unbreakable without the initialization vector. None of the stations on the ring will have access to both the encrypted message and the initialization vector except the transmitting and receiving stations. The generation of the initialization vector is unique for each confidential transmission and thus provides a unique approach to the key distribution problem. The FDDI protocol is of particular interest to the military in terms of LAN/MAN implementations. Both the Army and the Navy are considering the standard as the basis for future network systems. A simple and reliable security mechanism with the potential to support realtime communications is a necessary consideration in the implementation of these systems. The proposed method offers several advantages over traditional methods in terms of speed, reliability, and standardization.

For advanced fault tolerant satellite systems, new data bus architectures and media access protocols have to be developed to distribute common processing elements and functions among all subsystems. Different architectures with differing access protocols and topologies have to be traded to arrive at an optimum bus architecture for Satellite High Speed data transfer applications. In this paper, we discuss a High Speed Data Bus (HSDB) architecture and evaluate Media Access Control (MAC) protocols. We simulated Fiber Distributed Data Interface (FDDI) and Reservation Access Protocols applicable to HSDB. Performance Metric including throughput, delay, queue length and station latency are considered. We proposed a deferred priority protocol and compared throughput-delay performance with preemptive priority and normal priority for FDDI.

This paper describes a simple low cost FDDI network design that eliminates the need for expensive brouters (bridges or routers) to interconnect lower speed LANs to the dual ring FDDI backbone. The multiple servers attached to the dual ring backbone provide a highly fault tolerant parallel processing capability approaching that of a mainframe. This is dubbed `MegaServer Architecture.' Design of the total integrated network in MegaServer Environment is described.

The Fiber Channel is a general purpose point-to-point serial connection standard permitting a variety of transmission media, including both copper and fiber optic transmission options with bit rates ranging from 266 megabits per second up to 1065 megabits. A switching fabric is defined for the Fiber Channel to allow the interconnection of large numbers of nodes at high bandwidth. The Fiber Channel standard defines the implementation from the physical media and connectors up through switched and connectionless transport level services well suited for high performance data transfers. For the attachment of disk and tape storage subsystems, a communication model is required to define the command, data transfer, and response sequences that access the subsystem. The Small Computer System Interface (SCSI) provides an internationally standardized architectural model optimized for the attachment of storage subsystems and other intelligent devices to host computers. This paper describes the mapping of SCSI into Fiber Channel and describes the architectural advantages of SCSI as a model for communication across the Fiber Channel.