Browse Prior Art Database

Improved Means for Implementing Optical Interconnections for Parallel Processors

IP.com Disclosure Number: IPCOM000120221D
Original Publication Date: 1991-Apr-01
Included in the Prior Art Database: 2005-Apr-02
Document File: 3 page(s) / 116K

Publishing Venue

IBM

Related People

Green, PE: AUTHOR [+2]

Abstract

This article describes an efficient method for using a broadband communication network to interconnect parallel processors and memories. Parallel computer interconnections have traditionally been made by electrical means. Among the techniques commonly used are a bus, crossbar, and multilevel switching network. In all cases, the high- speed switching networks used for interconnections use conductors dedicated to carrying a single transmission at any given instant of time. In no high-speed switching network of this type does one conductor carry independent transmissions concurrently at the same instant of time, although this is done in some low-speed local-area networks.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Improved Means for Implementing Optical Interconnections for Parallel
Processors

      This article describes an efficient method for using a
broadband communication network to interconnect parallel processors
and memories.  Parallel computer interconnections have traditionally
been made by electrical means.  Among the techniques commonly used
are a bus, crossbar, and multilevel switching network.  In all cases,
the high- speed switching networks used for interconnections use
conductors dedicated to carrying a single transmission at any given
instant of time.  In no high-speed switching network of this type
does one conductor carry independent transmissions concurrently at
the same instant of time, although this is done in some low-speed
local-area networks.

      Optical fibers have extremely large bandwidths. By using
transmitters of different frequencies, and by tuning receivers to
specific frequencies, an optical system can carry multiple
conversations concurrently.  A star-coupler device joins many
point-to-point fibers to produce a multi-input, multi-output optical
interconnection, which is essentially equivalent to a bus in a
parallel computer system that has the additional property of being
able to broadcast all signals from all inputs to all outputs in the
same instant of time.  The connections may be made dynamically
switchable by providing frequency tunability in transmitters,
receivers or both.

      It has been demonstrated in the preceding article that an
optical interconnection system consisting of fibers, a star coupler,
tunable transmitters, and tunable receivers can be used as the
interconnection system of a parallel computer system.  The scheme
described there requires that at each processor and each memory unit
there be two transmitters and two receivers.  One processor
transmitter and one memory receiver tunes over N frequencies Fj in a
Group P of N frequencies in order to process the request and grant of
a connection, and then a second set tunes over N other frequencies in
a separate Group D in order to process the data transfer.

      Here we show a scheme that reduces the number of frequencies to
N.
HARDWARE DESCRIPTION

      The hardware consists of at least the following components:
1.  A star-coupler optical link that connects N inputs to N
    outputs,
2.  N independent processors,
3.  N memory modules,
4.  One tunable laser transmitter at each processor,
5.  One tunable receiver at each processor...