July 1988 Report on the ADVANCED COMPUTER COMMUNICATION WORKSHOP Lake Arrowhead, California March 30-31, 1987 CONTENTS Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Introduction - Danny Cohen, USC-ISI . . . . . . . . . . . . . . . . . 5 Opening Remarks - Gordon Bell, NSF . . . . . . . . . . . . . . . . . 9 Presentations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1. Integrated Services Digital Network (ISDN) . . . . . . . . . . 12 Daniel Sheinbein AT&T-BTL 2. Wideband Services . . . . . . . . . . . . . . . . . . . . . . 13 Ken Ingram AT&T-BTL 3. Multichannel Multihop Local Lightwave Network . . . . . . . . 16 A. S. Acampora AT&T-BTL USC-ISI Arrowhead Report [Page 1] Advanced Computer Communication Workshop July 1988 4. Fiber Distributed Data Interface (FDDI) and . . . . . . . . . 17 Using FDDI II to Mix Voice, Data, and Video Dono Van Mierop Fibronics, Inc. 5. Bitstream Processing System . . . . . . . . . . . . . . . . . 19 Eli Pasternak Telestream 6. Concept of Quanta Switching . . . . . . . . . . . . . . . . . 21 Brendan O'Dowd O'Dowd Research 7. Integrated Switching Over a High Speed Packet Network . . . . 23 W. David Sincoskie Bellcore 8. Current Research on Local Area and Wide Area Networks . . . . 27 and Processing Protocols at High Speeds A. G. Fraser AT&T-BTL 9. The Protocol Engine . . . . . . . . . . . . . . . . . . . . . 30 Gregg Chesson Silicon Graphics, Inc. 10. Blazenet: Fast Packet Switching . . . . . . . . . . . . . . . 33 David Cheriton Stanford University USC-ISI Arrowhead Report [Page 2] Advanced Computer Communication Workshop July 1988 11. NASA Science Internet (NSI) Overview . . . . . . . . . . . . . 36 Steven Goldstein NASA/Mitre 12. The FAST Project . . . . . . . . . . . . . . . . . . . . . . . 37 Jon Postel USC-ISI 13. Five Observations on Improving Future Data Communications . . 38 Steven Lukasik Northrop Corporation National Communication Initiative Issues - Danny Cohen, USC-ISI . . . 39 Discussion of NCI Issues - Larry Roberts, Net-Express . . . . . . . . 41 Summary of NCI Issues - Jon Postel, USC-ISI . . . . . . . . . . . . . 42 Closing Remarks - Gordon Bell, NSF . . . . . . . . . . . . . . . . . 45 List of Participants . . . . . . . . . . . . . . . . . . . . . . . . 47 USC-ISI Arrowhead Report [Page 3] Advanced Computer Communication Workshop July 1988 PREFACE This report is a summary of the presentations and discussions at the Workshop on Advanced Computer Communications, which was sponsored by the National Science Foundation and was held 30 and 31 March 1987 in Lake Arrowhead, California. As Gordon Bell said in his opening remarks, "the task of this conference will be to consider and attempt to define the requirements for a unified network strategy, along with the technology to implement such a strategy". A major point of discussion throughout was the relationship between communication as organized by computer data specialists (packet networks) and communication as organized by voice communication specialists (circuit networks). The future direction of computer communication was discussed in the context of the state of the art and emerging technologies and techniques. A transcript of the audio recording of the workshop was edited to create this report. Considerable effort was made to avoid erroneously attributing remarks to the various speakers; any errors that may remain are solely the responsibility of the editors. ACKNOWLEDGMENTS Thanks are due to the staff of the UCLA Lake Arrowhead Conference Center, to Jan Brooks of USC-ISI for acting as the conference coordinator, and to Sharon Dichter for making sense of the transcripts and doing the bulk of the editing of this report. In addition we thank all of the participants for the lively discussions and regret that this report does not capture them. We particularly thank Vint Cerf and Larry Roberts for being discussion leaders. USC-ISI Arrowhead Report [Page 4] Advanced Computer Communication Workshop July 1988 INTRODUCTION Danny Cohen USC-ISI Welcome to the first annual workshop on Advanced Computer Communication, sponsored by NSF and hosted by USC-ISI. We would like to express here our appreciation and thanks to NSF for sponsoring this workshop. We have invited to the workshop some elder statesman, representative of carriers and of equipment vendors, government people, and several computer network users (mostly from the academic research community). The purpose of the workshop is to explore issues about future computer communication networks. We plan to have in the workshop just a few presentations, and much discussion from which we can call learn. We are interested in future computer communication that is typified by: * Faster Data Rates (DS-0, DS-1, DS-3, FDDI and beyond) * Many Users (millions, universal, ubiquitous, open to all) * New Services (buying information, legally binding transactions, purchasing, banking, entertainment, education, directory assistance, yellow pages, workplace...) The notion of faster data rates, or higher performance, has more dimensions in addition to the bit/sec, such as low latency and (bit*mile)/sec, like the airline measure of passengers*miles. The ARPANET is a DS-0 based wide area network. The move to DS-1 based wide area networks has already been started and just as wide area networks got to be a small step behind the local area networks, fiber came along pushing local performance to the hundreds of Mb/s with Gb/s just around the corner, with wide area networks laging behind the local ones. We expect that future wide area networks will grow from the current DS-0 and DS-1 to DS-3 and to "fiber performance" of Gb/s, and some day even beyond that. We expect that the computer communication networks of the future will have millions of subscribers, similar in numbers to telephones. There are already over 150,000,000 telephones in the US alone, and about twice as many in the entire world. Therefore, we should expect a global computer communication network of several hundred million users, open to USC-ISI Arrowhead Report [Page 5] Advanced Computer Communication Workshop July 1988 all (research, private industry, homes, government, etc.), anywhere (internationally). Just as any telephone anywhere (at least in the more developed places) can call any other phone in the world (ditto), computers should be able to communicate universally, too. Communication networks can achieve such magnitude only if their users pay for the services. Unlike feasibility demonstrations that we routinely conduct in academia, such an ubiquitous communication network could exist if, and only if, it complies with the economic reality: it must be paid for and should charge for services provided. In our research world we marvel at new technology and ask what does it do. In the real world they look at it and ask how it is paid for. The kind of services that will be available in the future is limited only by our imagination (and/or lack of). We expect buying/selling of information, legally binding transaction (banking, purchasing,...), entertainment, education, and a workplace for millions of people. Multimedia communication will, so we believe, change fundamentally the way people communicate with each other over distances. There are several reasons why we use packet-switching. Packet-switching provides multi-destination connections with indefinite number of simultaneous parties ("parallel sessions"). Packet-switching provides automatic configuration of networks; packet-switching is an efficient on-demand use of the bandwidth through TDM. We had to develop the packet-switching technology because there was not at the time (late 60's) any way to buy these services from the carriers. At this point Larry Roberts added the following: We developed packet-switching because: * We needed better reliability than the poor reliability that was available then; * We needed variable bandwidth on demand (not just 2.4 Kb/s); * We needed to be able to rapidly change the connectivity, in a rate that was impossible by phone calls that took 30-60 seconds to establish; * We could not buy then what we needed at a reasonable cost. and Bob Kahn added the following: * We needed to obtain rapid switching (msec) of wideband circuits (56 Kb/s); USC-ISI Arrowhead Report [Page 6] Advanced Computer Communication Workshop July 1988 * at cost effective rates (linear in number of sites instead of quadratic); * and do it in finite time (months vs. decades). * Mini-computers made nodal switching practical. * The by-product was a quick alternate routing via packet- switching, error free communication, and asynchrony at host interfaces. Computer communication networks are separated from the phone networks. The $64,000 question is: Should they be separated? Another interesting questions is are they really separated? Since the ARPANET DS-0 links, for example, are part of the general switched telephone network (STN) it turns out that not only are the networks not separated, but also the ARPANET often makes suboptimal use of the STN. In the beginning the carriers ignored the computer communication networks, probably under advice of their lawyers, not of their engineers. As a result, computer communication networks today practically ignore the STN, except for buying dumb pipes through it. I believe that the carriers are interested in changing this. This change is not easy. I hope that it would start soon. The carriers model of the communication world consists of smart switches inside the STN and dumb terminals outside. Our model of the communication world is of dumb pipes through the STN and smart switches outside of it. Where should the smarts be? All inside the network? all outside? or appropriately divided among both? How do we get from here to there? Payment for services rendered is the key to the economic vitality of any real world systems. In Disneyland, for example, customers may pay for each ride according to its value by using A through E tickets (sort of "a la carte"), or they may pay only a single admission fee for unlimited use of all the rides (sort of "all you can eat"). Most young visitors have a third model of payment: "The Parents Pay". We, in the research world theoretically use the "pay once at gate" approach, but in reality we have grown accustomed to having our parent sponsors pay for all of our communication needs. As a result we use only one type of service, the equivalent of Special-Delivery telegrams. We must realize that our convenient economic model, "the parents pay", cannot be the basis for global communication systems for millions, or several hundred million users. USC-ISI Arrowhead Report [Page 7] Advanced Computer Communication Workshop July 1988 In order to communicate information, not just bits of data, we need protocols. There are several approaches (aka "architectures" and "models") to protocols, with our own Internet's IP/TCP and with the ISO's sacred seven layers for OSI leading the popularity contest. Some protocols are connectionless, others are connection-based, some value reliability above all, others prefer realtime capabilities. It seems (at least to me) that the existing popular protocol leave too much to be desired for operation in the future high performance, ubiquitous computer communication networks. If so, how shall they be fixed? or should pursue new approaches even if this would require us to abandon the existing protocols in favor of new protocols based on other approaches? In other words: shall we plan to necessarily evolve from the present protocols to the next generation ones, or should we be ready for some revolution. We use two basic communication paradigms: message passing and virtual circuits. Do they suffice? Should we look at other ones such as remote direct memory transfer? In summary, future computer communication networks present us with a great technical challenge that most likely could be met only through new approaches to computer communication. Personally, I believe, that it should be met through close cooperation between the computer communication research community and the carriers. I hope that this workshop would help in starting the convergence between these two currently diverging communities. USC-ISI Arrowhead Report [Page 8] Advanced Computer Communication Workshop July 1988 OPENING REMARKS Gordon Bell National Science Foundation One purpose of this workshop is to define the problems we are trying to solve. In February (1987), a conference was held in San Diego to produce a report to respond to the Gore Bill; this workshop has the advantage of having fewer participants. It may be necessary to have an even smaller group work on the problem definition, but the work done at this workshop can form a basis for future efforts. First, there must be a definition of the network itself and a projection of future network needs. We need to know what we are trying to do before we attempt to do it. For example, when the ARPANET was developed, electronic mail was not considered to be an important application. Now, 50% of its traffic appears to be mail. We must define the characteristics of a network which would link the research community together. The Gore Bill proposes that the government undertake a study of the critical current and future problems for communication networks for research computers at universities and federal research facilities, as well as for industry. The study would analyze the network needs for the next 15 years, including traffic, reliability, software compatibility, graphics, and security. In addition, the study would report on the benefits and facilities that must be offered by the network (something about which there is currently little information), and describe available options, including fiber optics. This would be in two phases. The first phase, due this summer, would address supercomputer access to the network, and the second phase, lasting about two years, would address all computers. Currently, the existing data network offers an unacceptable level of quality in terms of response time and reliability. Furthermore it is an operational nightmare. Judge Green's vendetta, the breakup of the AT&T operating companies, further complicated the problems. Line unavailability is a problem, and if switching is considered, it becomes evident that the system does not work. The ARPANET, 15 years old, is outgrown. One hundred computers using the 100 Kb/s switching capacity works out to one Kb/s each. When there are 1,000 computers using the same capacity, the result is 100 b/s each. This is what some users measure. All of these networks certainly require money. For example, CSnet has 200 sites and 10,000 people, and generates $1.5 million per year as a user-supported network. However, planning, rather than money, seems to USC-ISI Arrowhead Report [Page 9] Advanced Computer Communication Workshop July 1988 be the major problem. There are already 4,000 supercomputer users at 200 sites in the NSF, and 3,000 computers at 50 Department of Energy (DOE) sites and universities, with approximately 30,000 scientists. By 1995, DOE would like to be running 125 MB per day interactive with an interaction rate on the order of 20 Mb/s per person. Campus area networks, or CANs, would operate at 100 Mb/s. This indicates what might be the demand. The problem is that these networks are not interconnected. An analogy would be to have a phone on the desk (prior to Vail) for each network for which there is a connection. There could be five or ten phones on the desk, or, to extend the analogy further, multiple phones on multiple desks. The technical problems are primarily concerned with switching capacity. What we want to transmit determines the speed at which it can be transmitted. For example, video would require 5 Gb/s; supercomputers, 300 Gb/s; workstation graphics, 300 Gb/s; and loosely coupled workstation-to-host connections, 10 Gb/s when you have hundreds of sites. The required peak operating capacity for fiber networks would be about 10 Gb/s, and 1,000 Gb/s for a bundle. Currently, the capacity in a fiber network is 1.7 Gb/s. There are a number of fiber options, but the switches are not available. ISDN may be a part of the answer to the problem, but there is no time line for implementing it, and there is no certainty that the right questions are being asked. There are start-up companies currently working on these issues, but the scope of the question may be beyond them. A model more like the Federal highway system might be appropriate, where to build the infrastructure required the Federal government to pay for the interstate highway system, and the local and state highways were paid for per user. For data networks, some degree of user involvement in paying for network planning and implementation is required. The organization of such a network is not clear. Starting with 100 b/s and going to 10 Gb/s over a forty-year period, projections indicate that there will be considerable traffic with many links. Historically, with the ARPANET, there was a big increase in speed to something on the order of 10 Kb/s. With the current greatly increased use of the system, slower speeds of 600 b/s can frequently be measured. The ISDN 56 Kb/s service is starting operation in a few years. There is no network planned that is equal to T1 service in the same framework, even though protocols define what T1 service would be. In contrast, Japan is installing a slightly higher speed ISDN network at the T1 rate. However, any planning for a unified research net must take into account USC-ISI Arrowhead Report [Page 10] Advanced Computer Communication Workshop July 1988 that existing nets will be in use for some time. The only way existing nets would be subsumed is for a unified net to be put into place to replace them. Thus, it might be advisable to build a large backbone research net that could be the leading edge of this unification effort. The NSF has been considering the idea of regional networks. These are bottom-up nets being installed by various regions. There are four or five such networks initiated by the NSF that are already operational. Below this would be the consortia nets at universities and research centers. In summary, it appears that the task of this conference will be to consider and attempt to define the requirements for a unified network strategy, along with the technology required to implement such a strategy. The backbone net of supercomputers at a relatively small number of sites, meeting the needs of researchers at regional and campus area nets could serve as a model for consideration. We must address the question of user involvement, as well as the projection of capacity requirements through 1995. When we define what kind of network, with what capacity, and using which technology, it may be possible for us to address the question of how to fund the implementation and operation of the network. USC-ISI Arrowhead Report [Page 11] Advanced Computer Communication Workshop July 1988 1. "Integrated Services Digital Network (ISDN)" Daniel Sheinbein, AT&T-BTL The major feature of the Integrated Services Digital Network (ISDN) is its ability to allow premises and network processors to communicate efficiently with each other. This provides for greater functionality than is available with either premises or network processors alone. There are two ISDN interfaces which are currently defined: Basic Rate and Primary Rate. Basic Rate Interface (BRI) is a loop interface which supports 144 Kb/s and consists of two "B" channels of 64 Kb/s and a "D" channel of 16 Kb/s. The "D" channel supports out-of-band message signaling that utilizes a flexible message-based arrangement based on the Q.931 protocol. This protocol allows efficient processor-to- processor communication. BRI gives users substantial improvement over current loop capabilities and utilizes the existing copper wires connecting a user to a PBX or Central Office. Primary Rate Interface (PRI) is a T1 interface that supports 1.5 Mb/s and consists of 23 "B" channels and a "D" channel of 64 Kb/s each. The "D" channel supports out-of-band message signaling that utilizes a flexible message-based arrangement based on the Q.931 protocol. The PRI is intended for PBX-to-PBX, PBX-to-Network, Host-to-PBX, or Host-to-Host network connection. ISDN provides an evolutionary set of data networking capabilities. Circuit switched data, in which the data are sent on the "B" channel will be supported, as will access to X.25 packet-switched services. In addition, the ISDN Packet Mode (or ISDN Frame Relay) is being defined. This will allow the rapid transmission of data with low latency. The new key element of ISDN Frame Relay is that transmission error recovery and flow control responsibility have been removed from the network and are the responsibility of the end system. This capability, in addition to the inherent flexibility of packet-switching and virtual circuits, permits the elimination of private lines. The ISDN Network also provides flexible routing on demand. The natural extension of ISDN Frame Relay is a non-channelized version that will allow the user to employ any increment of capacity and will not be limited to 64 Kb/s chunks. ISDN is not static. Work has begun on the definition of a "Broadband" ISDN (B-ISDN). B-ISDN is expected to be defined in the 140 Mb/s range and will have user-selected channelized and non-channelized bundles, utilize the frame relay approach, and will use fast packet technology for implementation. USC-ISI Arrowhead Report [Page 12] Advanced Computer Communication Workshop July 1988 2. "Wideband Services" Ken Ingram, AT&T-BTL This presentation discusses AT&T-BTL'S current and future plans for wideband facilities. The current telephone system is a mixture of fiber, digital radio, and coaxial cable. The coaxial system was installed coast-to-coast, and some of the trunks on the system have been converted to digital. The characteristics of digital radio have changed significantly. When today's forward error-correcting techniques for digital transmission are coupled with non-disruptive protection switching, manual protection switching, and new antenna applications, the same quality of digital transmission is available on radio as on fiber. (That's certainly not been true in the recent past when modem technology was used to provide digital facilities over routes engineered for analog radio.) Typically, in the future, fiber will be used for high-demand, high- growth areas and bandwidth requirements greater than that provided with radio. Radio will be for the lower demand, lower growth areas. The current deployment for fiber is 417 Mb/s, which is nine DS-3s. There are generally seven working systems and one protection system in the cross section of fibers. The centralized operating support system gathers data on the "health" of each regenerator and the status of every line in the fiber systems. It tells what lines are in use, what lines are turned down, and what lines are available for restoration. The present system has 25-mile nominal repeaters and up to 180 Mb/s wave division multiplexing (WDM). This allows two systems to be put on a given fiber pair by using different wavelength repeaters. (In two places in the United States, WDM of 417 Mb/s is being deployed.) The typical bit error rate performance of this system is better than 10**- 11. Presently, the backbone switch for the network is the 4-ESS, but a number of 5-ESS switches are being deployed in areas where switching volumes are smaller. In addition, non-hierarchical control routing is being deployed to get away from old hierarchical routing (which is always upchain). When that is completed, there will be no more than two circuits switched in tandem to complete any call, but any particular call could have 27 potential routes. When the system is fully deployed, only two hops will be necessary to complete a call, and that will reduce overall circuit switching time. The basic advantages of very fast packet-switching are being considered. This will give the ability to be able to control, manage, provision, and USC-ISI Arrowhead Report [Page 13] Advanced Computer Communication Workshop July 1988 troubleshoot any virtual circuit in the same way, providing major administrative and maintenance procedure improvements. In the near future, 1.7 Gb/s systems will be deployed. In fact, by the middle of 1988, most of the single mode routes will be redeployed as 1.7 Gb/s systems. If 1.5 micron lasers are used, economic balance will be provided by allowing a longer repeater spacing, instead of using the shorter repeater spacing and the 1.3 Gb/s and 1.5 micron systems in the WDM arrangement. Development has already been funded for 1.7 Gb/s WDM since needs have been identified for that in some of the routes in early 1990's. This will provide a fairly significant amount of available bandwidth, giving some interest in the possibility of 400 Mb/s transmission rates across cross sections. In addition, a multi-gigabit bandwidth (somewhere in the 7 Gb/s range) field operating experiment is being considered for the early 1990's. Various composites of fibers that have potential for 100-mile-plus repeater spacing are also being considered. This is probably going to be more valuable in the undersea cable world where repeaters under the ocean are not wanted. Direct optical stimulation of modems and computers and direct optical switching are also being looked at. Coherent systems will have almost unlimited systems capacity if tunable lasers and receivers become available, so that fiber pairs can be treated as if they are radio waves. In this case, understanding bandwidth data transmission requirements will be an advantage to all users involved. For example, when 1.5 Mb/s (T1) was implemented and bandwidths were provided in a non-channelized atmosphere, it was learned that special things had to be done to perform end-to-end error checking and trouble isolation. Because that understanding did not exist in the early days of the 1.5 Mb/s (T1) network, a tariff structure for T45 does not exist yet. It is necessary to have the proper operating characteristics and the ability to shoot and clear trouble quickly before the tariff structure is developed. Also, it is not yet clear what multiple of T45 might be used; a good understanding of the wideband data transmission requirements will provide the flexibility necessary for any of those multiples to be implemented. By the end of 1989, another transcontinental fiber route will be added. More important, in the denser areas of population a number of natural rings will be formed. The alternate routing capability coupled with the multi-gigabit transmission capability will quadruple the capacity of any given route at the same price it will cost to double it using WDM. This very economical method coupled with some sensing, and perhaps even optical switching at the line rate, will be able to perform automatic detection and restoration of total fiber route failures. USC-ISI Arrowhead Report [Page 14] Advanced Computer Communication Workshop July 1988 If large capacity systems exist with current restoration capabilities, a significant impact on the quality of service will occur, unless today's technology can be applied. For example, if current technology, which senses a single line failure and switches it within milliseconds, is enhanced with some network mapping information, so that the status of each potential alternate route is known, a total route failure should be restored in about 10 seconds. This will result in some very significant changes in the basic reliability of the backbone network itself. USC-ISI Arrowhead Report [Page 15] Advanced Computer Communication Workshop July 1988 3. "Multichannel Multihop Local Lightwave Network" A. S. Acampora, AT&T-BTL Because the speed at which light may be electro-optically modulated is limited to a several Gb/s range, a network where all users access a single channel provides an overall capacity that does not tap the vast bandwidth potential of lightwave technology. A fundamental feature to seek in a network is the ability to provide an aggregate network capacity shared by all users and far in excess of the peak rate at which any user may access the network. One shortcoming of this network is that the entire band is not instantaneously available to all users and connectivity among users becomes a concern. To date, proposed connectivity techniques require rapidly agile or tunable active optical components to enable communications among transmit/receive user pairs in response to the instantaneously changing bursty traffic demand. Such optical components do not presently exist. The Multichannel Multihop Local Lightwave Network is a new architecture that is particularly well-suited for lightwave implementation. Using only passive optical components, this architecture vastly multiplies the capacity of a single-channel network by assigning two fixed wavelength transmitters and receivers to each user. The approach physically uses a distributed bus, star, or tree topology and several independent channels that are wavelength multiplexed onto the fiber transmission medium. By relaying through intermediate nodes, wavelengths are assigned to users in such a way that a connection may be established between any pairs of users assigned to different transmit/receive wavelengths. The achievable capacity grows monotonically as more users and channels are added to the network. Typical results for an access speed of 1 Gb/s are that a network with about 1,000 users can provide an aggregate capacity of approximately 200 Gb/s. Thus, each of 1,000 users can transmit at a peak rate of 1 Gb/s and an average rate of 200 Mb/s. The queuing delay is negligible for offered loads equal to 80% of the achievable capacity. For that loading, the lost packet rate caused by buffer overflow is less than 1 in 106 if each user's queueing buffers can store 128 packets. The advantages of this approach are that it is modular and reliable, it requires neither centralized control nor coordination among users, and it employs high-speed electronics vastly more efficiently than a single-channel network. USC-ISI Arrowhead Report [Page 16] Advanced Computer Communication Workshop July 1988 4. "Fiber Distributed Data Interface (FDDI)" Dono Van Mierop, Fibronics, Inc. Fiber Distributed Data Interface (FDDI) is the standard currently available in high-speed networking. It is suitable for a wide range of applications including backbone and backend networks and workstation connections. One of the future applications may include integrated services networks. The development of the FDDI standard is based upon the work of the American National Standards Institute (ANSI) Committee X3T9.5. and relates to the Physical and most of the Data Link layer of the Open System Interconnect (OSI) model. The major characteristics of FDDI and FDDI networks are: o FDDI architecture is modified from token-ring architecture, which grew out of the IEEE 802.5 token-ring standard. To ensure high speed, high reliability, and availability, Media Access Control (MAC) remains a token that passes around the ring; whoever captures the token may transmit, and must release the token after transmission. The duration that the token can be kept is subject to restrictions designed to assure fairness to each user. o An FDDI network supports a range of ring sizes from small to large and utilizes ring (physical) or star (using concentrators) topologies. It performs optimally for both large and small rings. o FDDI is based upon a fiber optic medium (125 micron outer diameter) with duplex connections and electro-optical bypass switches. The wavelength is 1.3 micron. The medium is insensitive to sources of electromagnetic interference and the fibers do not emit electrical noise, thereby providing a low bit error rate (10**-9). o An FDDI network is unlimited in its length and the number of stations that can be attached to it. Typical networks range from a few stations to hundreds of connections, and stretch from a few hundred meters (computer room backend network) to hundreds of kilometers (a metropolitan area application). o The performance of an FDDI network is not sensitive to the size, number of stations, or loading of the network. o The distance between two stations on an FDDI network is restricted to 2 km so that relatively inexpensive Light Emitting Diode (LED) technology may be employed. o Adjacent stations on an FDDI network are connected over duplex fiber USC-ISI Arrowhead Report [Page 17] Advanced Computer Communication Workshop July 1988 which creates a dual ring used in two ways: to double the throughput; and to provide redundancy, which serves as the basis for automatic recovery from failures. o The raw speed of the FDDI network is 100 Mb/s with the actual bit rate being 125 Mb/s, due to unique encoding. The sustained data rate is well above 80 Mb/s because of inherently high efficiency. This is about one to two orders of magnitude higher than standard LANs. Fibronics, Inc. has developed a first implementation of FDDI-based products called System FINEX. The system is composed of a set of FINEX Controllers (FCs) interconnected via an FDDI-based fiber optic cable plant. Each FC is connected to one or more devices, that is, LANs, hosts, or workstations. Tests of the system have been encouraging and development of FDDI-based products for constructing backbones, backend networks, and workstation interconnections continues. "Using FDDI II to Mix Voice, Video, and Data" FDDI II, a superset of FDDI I, includes circuit switching, in addition to packet-switching capability. This provides operations for time- sensitive data such as voice, video, etc. The concept of FDDI II is simple. A multiplexer is placed between the physical and media access layers and multiplexes traffic through as either packet or virtual circuit traffic, providing a virtual circuit server that goes through multiple networks. The multiplexer uses a time-division technique. A master station sends a frame every 125 microseconds. Each frame includes a preamble that determines whether the service is packet or circuit switching. Each channel is 6.144 Mb X 16 or almost 100 Mb. 6.144 Mb is three times the European T1 and four times the American T1. The FDDI II standard is currently being worked on by a special subcommittee, and there are vendors working on chips that will probably be available in the next two to three years. USC-ISI Arrowhead Report [Page 18] Advanced Computer Communication Workshop July 1988 5. "Bitstream Processing System" Eli Pasternak, Telestream System architecture used for high-speed communications applications has three fundamental requirements: high data throughput (on the order of 1 Gb/s); large amounts of bitstream processing power (on the order of 100 MIPS); and reasonable cost. Implementation of such a system with open architecture and a significant amount of programmability are also desirable, as they allow use of proprietary interfaces and protocols and provide for the evolution of communication standards. The Bitstream Processing System (BPS), developed by Telestream Corporation, is a high-performance, programmable communications processor. It is capable of terminating, switching, and processing the content of several multi-megabit per second digital bitstreams, such as those existing in LANs and T1/T3 -based WANs. Using the BPS as a switching and processing platform, implementation of a broad range of tele- and data communications products (such as hybrid circuit and packet switches and LAN and WAN gateway processors) is feasible. The BPS has been designed to support the high speeds and deterministic response times associated with circuit and fast packet-switching, as well as the software-intensive, event-driven processing associated with the management of the data flow in a network. A 1.28 Gb/s time- division, multiplexed bus -- the heart of the BPS -- provides intertask communication with no-wait states and interrupt support for event-driven operation. Processing and Function Elements perform content processing of multi- megabit per second bitstreams. Each Processing Element provides 10 MIPS of logical bit/byte processing power in a deterministic mode of operation. This results in guaranteed response times for time-critical processing, such as front-end circuit and packet-switched processing. It is possible to configure up to 16 Processing Elements in parallel, resulting in 160 MIPS of bit/byte processing power per BPS shelf. Traditional microprocessors (such as Motorola 680X0 or Intel 286/386) implemented as Function Elements can perform non-deterministic or memory-intensive processing. Using appropriate VLSI or dedicated logic, Function Elements can also perform hardware-intensive processing, such as voice compression or encryption. The parallel architecture of BPS provides high-speed, bit/byte manipulation. VLSI integrated circuits provide low-cost bus access on each card. The BPS architecture supports both software- and hardware- intensive processing in similar manners, and functions can be easily transferred from software on Processing Elements to hardware on Function Elements as silicon becomes available. USC-ISI Arrowhead Report [Page 19] Advanced Computer Communication Workshop July 1988 Telestream provides software and hardware development tools and bus access to assist in the implementation of custom software and hardware for the BPS. In addition, BPS can be used to implement all three phases of the NSFNet; phases one and two require a single 19 inch shelf configuration, and phase three requires multiple shelves. USC-ISI Arrowhead Report [Page 20] Advanced Computer Communication Workshop July 1988 6. "Concept of Quanta Switching" Brendan O'Dowd, O'Dowd Research Quanta Switching, developed by O'Dowd Research, Inc. and its parent company, O'Dowd Research Pty., Ltd., of Australia, is a modular, high- speed communications architecture allowing universal communications networks to be implemented that provide: o efficient and reliable integration of voice, video, and data communications over a common network o instantaneous communications paths in the hundreds of Mb/s, regardless of distance, for uncompressed video, Computer-Aided Design (CAD)/ Computer-Assisted Mechanization (CAM), LAN, and/or bulk file-transfer applications o inherent protocol conversion between dissimilar terminal and host environments o resource enhancement for data terminals through per-device intelligence o compression of voice and data communications for the most efficient utilization of costly transmission facilities. These network attributes are realized through a set of switching, gateway, and terminal adaptor components based upon the fundamental architectural building block of the Quanta unit of information. A brief summary follows: o The Quanta is a fixed-length, 192-bit information packet, containing type, address, information, and error-checking fields. Packets are considered to be completely asynchronous in nature and there are two basic types: Command Quanta and Information Quanta. For most applications, packets are assembled in small, low-cost, per-device Quanta Network Adaptors (QNAs). o A Quanta network allows significantly enhanced resources through the per-device intelligence residing in the QNAs placed in or at each terminal. In addition to assembling information into Quanta units, each QNA is responsible for device, call, and information transfer management. Data terminal QNAs provide both protocol conversion from the terminal-specific protocol to a universal protocol and multiple session capability for each terminal. After the initial session has been established, all QNA management USC-ISI Arrowhead Report [Page 21] Advanced Computer Communication Workshop July 1988 of information transfer is handled in hardware. This provides low- cost communications, and frees up processing power that can be used to provide additional functions such as protocol conversion and multiple session support. o A Quanta is transmitted to a local or remote Quanta Communications Exchange (QCX) for network routing after being generated by a QNA. A QCX has up to 256 Quanta Interface Ports (QIPs) that are interconnected together via a parallel loop backplane. Each QCX QIP port supports up to 64,000 virtual channels and the match between virtual channels and physical end-points is selectable. All QCX QIPs on the parallel loop simultaneously transmit and receive Quanta to and from the loop, giving the Quanta loop a capacity at least as great as the underlying medium. Implementing all switching functions, except for call set-up, in hardware achieves high-speed switching and very low transit delays. State machines are used to recognize addresses, error conditions, priority levels, congestion, etc. o Quanta Gateways are implemented at each interswitch trunk interface to create a multiswitch network. The gateways perform address translations between trunk virtual circuits and handle all terminal error conditions. QCX Gateways interface to normal transmission lines, such as 64 Kb/s ISDN lines, 1.544 Mb/s T1 trunks, and 45 Mb/s fiber links. Special gateways are available to interface to external X.25, Ethernet, and other networks. o A Quanta network accommodates two basic types of information flow: data traffic and voice traffic. Because of time sensitivity, voice traffic is given priority over data traffic within the network. Bandwidth allocation procedures during call set-up ensure minimal voice Quanta delay and loss. Data traffic receives its own allocation of bandwidth and uses the extra bandwidth provided by instantaneous lulls in conversation. This ensures the most efficient use of transmission facilities and gives each end-user the highest possible level of performance. USC-ISI Arrowhead Report [Page 22] Advanced Computer Communication Workshop July 1988 7. "Integrated Switching Over a High Speed Packet Network" W. David Sincoskie, Bellcore It is necessary for the current network to evolve into a network that is extensible, flexible, and has an architecture usable for many different applications. This is preferable to a network that is optimized, but has only one application. Packet-switching is the only way known to implement flexible bandwidths, but current packet switches are slow (1,000 pkt/s and throughput of 1 Mb/s) and require complex protocols and general purpose computers. Broadband packet switches that are fast (1 to 10 million pkt/s and throughput of 10 Gb/s), use simple protocols, and are implemented in hardware (VLSI) are currently being developed. There are two ways to build packet switches. One way is to use a cross-point, N2, not self-routing type of fabric through which the packet is switched. Tony Acampora, of AT&T, is working on that research. The second way is to build a packet switch using a Banyan-Batcher network. This method grew out of the Starlight work done by AT&T, slightly before divestiture. A Banyan network can block internally for some set of inputs, so congestion in the switch occurs. This network also has the interesting property that if inputs are sorted in advance, the network becomes internally non-blocking. Research has resulted in the creation of some divergent ways to build Banyan networks. One is to put in internal buffering to handle congestion. This is the type of switch that AT&T used in their San Francisco Bay Area trial, and the switch that John Turner proposes. Another set of variations have been developed. Some have been built and others have just been analyzed. One is the combination of a Batcher and a Banyan network. A Batcher network is a N log2 N growth network. It takes a set of inputs and sorts them. The output is run into a Banyan network, an N log N growth network, which expands the outputs and delivers them to the appropriate ports. This uses a bit-serial line. It is hoped to get the data rate up to 40 Mb/s. The advantages of the Batcher-Banyan network are that it is bit rate independent, fully integrated, and non-blocking; has distributed control, high capacity, and fixed latency; and uses minimal hardware. USC-ISI Arrowhead Report [Page 23] Advanced Computer Communication Workshop July 1988 The disadvantages are that there is tight synchronization throughout the entire network, and there is a non-redundant path. All of the bits have to hit the stages simultaneously in order for the switch to be non- blocking. In the lab at Morristown, N.J., a 32-byte, 32 Batcher at a 32 X 32 merge (which allows the building of bigger Batchers) and a 32 X 32 Banyan have been built into a small packet switch. This switch is operational and consists of eight ports running at 55 Mb/s per port. It is felt that it is possible to get up to 256 parallel inputs, each one taking a serial packet, running to 100 Mb/s in 2.5 micron CMOS. Higher speeds will result if lower feature-size CMOS or gallium aresenide sites are used. The result will be 100 Mb/s plus or minus 50%, giving a switch with a raw bandwidth of approximately 18 Gb/s. One hundred percent occupancy is not possible unless an ideal traffic stream exists. If the input stream is 60%, a 100 Gb/s switch would result. The fabric for the switch would be packaged in a box of approximately 2 cubic feet and use 528 CMOS chips. The buffers can be done several different ways: at the output internally in the fabric, or at the inputs. There are different performance trade-offs for each of these. Due to several factors, this network has limited growth. First, there is synchronization required at each stage, as the bits have to hit each stage simultaneously. Secondly, problems result from wire interconnection and the variance in the length of the wires. As this network expands to 256 ports, a very large variance in a planar structure appears. One solution developed for the wire problem is a three-dimensional packaging structure. The alternate stages of the network are turned 90 degrees. This provides a 256 X 256 Banyan with a maximum length of four inches. This gets below the point where, at a 140 Mb, the interconnects become transmission lines instead of electrical wires. If virtual circuits are implemented through this network, the fabric can be run at 100%, because there are no output collisions. Sequencing is preserved throughout the network. Another problem that exists is the end-to-end routing problem throughout the entire network. Traditionally, this has been solved with a virtual circuit type arrangement. A packet carries the virtual circuit number, and the routing table routes the virtual circuit number to the output port. A simple table lookup provides end-to-end virtual circuit routing. An alternative is datagram routing. This type of routing addresses the USC-ISI Arrowhead Report [Page 24] Advanced Computer Communication Workshop July 1988 problem of both short-term and long-term mobility. The basic routing algorithm developed is an extension of the automatic, self-learning bridge routing algorithms that are used in the DEC LAN Bridge 100. By observing terminals emitting packets with their source addresses (very much like Ethernet addresses), the network learns where to route subsequent packets that are addressed to someone who has already emitted a packet into the network. A distributed datagram routing table can be set-up through network. One problem with this routing method is that the basic algorithm works only on a network without loops. A spanning tree structure is needed, but a tree does not have much capacity and is not reliable. Making these algorithms work in a network with 100 million hosts is necessary. A multiple-tree bridge-routing algorithm, which has a graph decomposed into several spanning trees, has been developed. The basic algorithm is applied to each individual tree, and a packet is marked so it goes on the tree. This allows the building of a complex network with an arbitrary capacity. Any graph can be partitioned into a set of spanning trees, then it can be routed. To apply this algorithm to today's telephone routing techniques, the routes are worked out in advance. The spanning trees are defined by going through a list of trees and choosing those trees with minimal error potential. A problem with extending the tree networks is that when the location of the host is unknown (that is, when it has been moved or disconnected), it must be located by broadcasting throughout the entire network, causing extraneous traffic. Based on a call rate of 100,000 calls per second across the U.S. in the busiest hour, the use of this basic algorithm might result in 100,000 pkt/s background traffic on the LAN. This is not good. An extension to the algorithm, called the multicast routing algorithm, allows for successive broadcasting in larger network chunks. One hundred thousand pkt/s of background traffic can be reduced to 1,000 pkt/s -- good for a fast packet-switching network. For a network consisting of a set of packet switches with trunks coming into them, there is a T-gate device that looks at the first address carried by each packet. It then tells the switch to route the packet to specific internal ports. The T-gate device that does the multicast algorithm has a database, which is a set of two holds. Each hold contains a destination address, a timer (used for timing out so the list is not clogged), and a bit vector of trunk numbers. Turning on a bit indicates that a copy of the USC-ISI Arrowhead Report [Page 25] Advanced Computer Communication Workshop July 1988 packet should be sent through the trunk.