Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Means of Implementing a Parallel Processing Barrier With an Optical Interconnection Network

IP.com Disclosure Number: IPCOM000101047D
Original Publication Date: 1990-Jun-01
Included in the Prior Art Database: 2005-Mar-16
Document File: 3 page(s) / 122K

Publishing Venue

IBM

Related People

Green, PE, Jr: AUTHOR [+2]

Abstract

Disclosed is a technique for utilizing the broadcast properties of optical connections to implement a barrier synchronization operation for parallel computers.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 50% of the total text.

Means of Implementing a Parallel Processing Barrier With an Optical Interconnection Network

       Disclosed is a technique for utilizing the broadcast
properties of optical connections to implement a barrier
synchronization operation for parallel computers.

      Barrier synchronization of multiple processors is an extremely
costly synchronization method when processors share a common bus.
When N processors have to synchronize at a barrier, at least N bus
transactions must take place. This is a costly overhead in a
high-speed parallel computer system.

      The broadband optical interconnection system disclosed here
provides a means for synchronization to be done concurrently with
normal traffic.  Consequently, such an interconnection system offers
the opportunity for higher performance than is available from prior
art that uses electrical connections on metal conductors.

      The prior art for electrical connections includes the use of
separate interconnections for the purposes of barrier synchronization
(*), and the use of an enhanced bus protocol to reduce, but not
eliminate, the overhead caused by barrier traffic.  The optical
interconnect provides a means for barrier synchronization to be done
within the same interconnection system that supports normal traffic
without requiring exclusive use of the interconnections for the
purpose of synchronization.

      The invention consists of an optical interconnection system
plus a protocol that implements the barrier function. The necessary
hardware within the optical interconnection system is the following:
1.  a star-coupler optical link that connects N inputs to N outputs,
2.  N independent processors,
3.  two laser transmitters per processor, one transmitting at gactive
and the other transmitting at gidle, and
4.  two receivers per processor, each capable of detecting optical
energy at one wavelength, which is gactive and gidle, respectively,
for the two receivers.

      A barrier in a parallel processing program is a point in the
program at which active processors must wait until all processors
reach the barrier.  When the barrier is reached, then all processors
or any subset of the processors can proceed with subsequent
execution.  Because programs can loop backwards and reach a barrier
multiple times, it is essential that no processor reach a barrier for
a second time until every processor that reached the barrier the
first time has observed  that the barrier condition has been met.  In
other words, no processor can check in at a barrier for iteration i +
1 until all processors that reach the barrier at iteration i have
checked out of the barrier.

                            (Image Omitted)

      To implement this function, we associate the two wavelengths
gidle and gactive with a barrier.  Each processor whose activity is
controlled by the barrier produces a transmission either at
wavelength gidle or at wavelength...