Browse Prior Art Database

Distributed Crosspoint Switch Control in a Parallel Processor System

IP.com Disclosure Number: IPCOM000105378D
Original Publication Date: 1993-Jul-01
Included in the Prior Art Database: 2005-Mar-19
Document File: 4 page(s) / 123K

Publishing Venue

IBM

Related People

Blum, A: AUTHOR [+3]

Abstract

A parallel processor system is described in which the control function and the supervisor function of a crosspoint switch are distributed.

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

Distributed Crosspoint Switch Control in a Parallel Processor System

      A parallel processor system is described in which the control
function and the supervisor function of a crosspoint switch are
distributed.

      For parallel processing involving a great number of satellite
processors, a crosspoint switch is a highly effective communication
means which needs a central supervisor function to control the con
nection and disconnection of communication links between the units ex
changing data.

      Normally, the attached satellite processors send connect
requests to this supervisor function if they require a link to one of
the re maining satellite processors for exchanging data.  The
requests are granted and the requesting satellite sends the addresses
of the source and destination satellites to the supervisor function
which checks a list in which all links are stored in order to obtain
the status of the addressed destination satellite processor.  If the
requested satellite is available, the link will be established and
the requesting satellite processor be informed.  The source satellite
then starts the data exchange.  At the end of the communication, the
supervisor function is reinvoked for the disconnect request by one of
the communicating sa tellites.  After receiving the satellite
addresses, the supervisor disconnects the link and updates the link
list containing the units that are busy communicating.

      This sequence involves communication overheads for the request
to travel to the supervisor, for the list management and the feedback
to the requesting unit before an operation to the destination
satellite can be started.  In order to avoid these overheads in terms
of time, logic circuitry, memory space, etc., it is proposed to
depart from the prin ciple of a centralized supervisor function and
to use the described distributed control function instead.

      Fig. 1 shows two of n satellite processors and their connection
by a crosspoint switch.  All n satellites and a switch controller are
interconnected by a dedicated bus (S-bus) transporting control in
formation for connect and disconnect requests.

      It is assumed that satellite A has been requested to
communicate with satellite B.  This is initiated by putting an
instruction into the mailbox of satellite A.  All actions following
now are assumed to be hardware-controlled.

      The BSA (Blue Bus to Switch Adapter) function (Fig. 1) then
reads the mailbox with the command, the satellite B address and the
data (or, depending upon the instruction, a storage address for the
dat...