Danny Cohen

Myrinet: a gigabit-per-second local area network

The Myrinet local area network employs the same technology used for packet communication and switching within massively parallel processors. In realizing this distributed MPP network, we developed specialized communication channels, cut-through switches, host interfaces, and software. To our knowledge, Myrinet demonstrates the highest performance per unit cost of any current LAN. >

On Holy Wars and a Plea for Peace

Which bit should travel first? The bit from the big end or the bit from the little end? Can a war between Big Endians and Little Endians be avoided?

Issues in transnet packetized voice communication

In recent years, very important progress was made in both real-time digital computer communication and real-time digital voice communication, due to the growing importance of geographically distributed computation, the growing need for resource sharing, and the need for secure voice communication. The progress in digital voice communication—mainly aimed at low data rate and high fidelity speech communication—has led to various vocoding techniques such as LPC and CVSD. The progress in real-time computer communication—which is mainly aimed at low delays, high bandwidth and high cost-effectiveness—led to technologies such as packet switching, packet-radio, and satellite communication, as well as intercomputer communication protocols. It is the natural result of the progress in both areas that the use of packet switching technology be considered for voice communication. In this paper the issues involved in the application of packet-switching networks to real-time voice communication are identified and explained.

Atomic: A High-Speed Local Communication Architecture

ATOMIC is an inexpensive high-speed LAN built by USC/ISI. It is based upon Mosaic technology developed for fine-grain, message-passing, massively paraIiel computation. Each Mosaic processor is capable of routing variable length packets, while providing added value through simultaneous computing and buffering. ATOMIC adds a general routing capability to the native Mosaic wormhole routing through store-and-forward and path concatenation. ATOMIC aggregate bandwidth scales as the number of nodes increases, and it has a small interface cost. Each ATOMIC channel has a data carrying capacity of 480 Mb/s. A prototype ATOMIC LAN has been constructed along with host interfaces and software that provides full TCP/IP compatibility. Using ATOMIC, 1500-byte packets have been exchanged between hosts at an aggregate transfer rate of more than 1 Gb/s. Other tests have demonstrated throughput of 5.25 million packets per second over a single Mosaic channel. The architecture is flexible both in topology and functionality. This paper describes the architecture and performance of ATOMIC.

The use of message-based multicomputer components to construct gigabit networks

The typical node of a message-based multicomputer consists of a microprocessor, router and memory. At the California Institute of Technology, the Mosaic project has integrated such a node onto a single chip. That reduction in scale fundamentally changes the scope of node application, since nodes become both very small, and inexpensive.Mosaic nodes may be employed to process, to generate, or to receive data. Since the router in a Mosaic node is independent of the microprocessor, computation and routing take place simultaneously. These nodes may be used to create general purpose gigabit LANs. They may also be used to create special purpose gigabit networks to interconnect instrumentation within spacecraft or aircraft.The ATOMIC project at USC/ISI is using Mosaic nodes to prototype a gigabit LAN testbed. This testbed is operational. Networking and administration software provides full TCP/IP compatibility. Packets have been exchanged between two interfaces at a rate above one gigabit per second (Gb/s).An individual ATOMIC interface is both inexpensive and small, consisting of one Mosaic chip, four SRAM chips and clock logic. Two interfaces easily fit onto a postcard-sized circuit board. Their low cost makes it practical to include several interfaces within a host, providing an interior Gb/s distribution network, multiple access points to the LAN for greater performance or redundancy, and other capabilities that are not yet fully explored. The results reported in this paper represent actual data obtained from the prototype.

Architectural and Software Issues in the Design and Application of Peripheral Array Processors

How can we overcome the limits of von Neumann architecture and avoid supercomputer costs? This overview suggests that peripheral array processors are the answer in many applications.

A protocol for packet-switching voice communication

Abstract This paper discusses the issues associated with real-time voice communication over packet switched networks. It suggests an approach for the design of a protocol to support this application. The ARPA Network Voice Protocol is presented as an example of such a protocol. In addition, two extensions to it are discussed. Most of the issues which are discussed in the context of the NVP, like separation of control from data, device independence, and performance monitoring are not unique only to voice application, but apply equally to any other real-time protocol. The major virtue of the real-time voice communication as far as networking is concerned is the real-time, not the voice, aspect of this communication.

Packet communication of online speech

The rapid progress of both communication and processing technologies in recent years constitutes a challenging technological revolution, epitomized by the photographs sent by Voyager from Saturn. This revolution offers new opportunities for many avenues of technology. Real-time speech communication is one of them. This paper discusses the use of packet switching, the emerging computer communication technology, for online speech application. The paper reviews the computer communication technologies such as packet and circuit switching, datagrams, and virtual circuits. It mentions the voice encoding (vocoding) algorithms and their implementation and discusses the interaction of vocoding and packet-switching technology. It argues about the issues of network voice protocols and introduces applications of online speech. After describing the experience with packet speech, the paper concludes with several ideas about the future of the field.

Mutual encapsulation of internetwork protocols

Abstract Encapsulation is a key concept in developing a layered architecture for communication protocols: objects from one layer can be “encapsulated” for transmission through a lower layer in the architecture. This approach can be used to combine different incompatible networks into a single “internetwork”: packets from a unified internet protocol are encapsulated within the network-specific formats and protocols associated with each individual network. But just as there are incompatible networks, we are now seeing the emergence of incompatible internetworks — systems serving different user communities, yet perhaps overlapping in their use of particular networks. Full inter-operability of these systems would require some form of protocol translation, or else the development of an even higher level “inter-internetwork” protocol. Yet different internetwork designs can combine some of their capabilities by invoking a higher level of encapsulation: each internetwork extends its range by using the other internetwork as one of its underlying communications links. This approach has been called “mutual encapsulation.” In this paper we explore the notion of mutual encapsulation; as an example we consider the coexistence of the Xerox internet environment with the Arpa internet environment. Each of these systems is composed of several individual networks with a variety of protocols and performance parameters; but with mutual encapsulation each can make use of the capabilities of the other.

A methodology for programming a pipeline array processor

In this note a recursive filter implementation is discussed, analyzed and programmed in the most efficient way for the FPS-AP120B array processor. The purpose of this note is not only to demonstrate a good technique for programming this filter (even though it is a noble goal by itself), but to demonstrate the methodology involved. The FPS-AP120B array processor was chosen because it has microprogrammed control parallel pipeline architecture, which is typical for a wide class of high performance array processors.