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

Purging Shared Cache Lines in a Multi-Processor Node

IP.com Disclosure Number: IPCOM000106388D
Original Publication Date: 1993-Nov-01
Included in the Prior Art Database: 2005-Mar-21
Document File: 4 page(s) / 95K

Publishing Venue

IBM

Related People

Baldus, DF: AUTHOR [+5]

Abstract

The destination field of a request in our multi-node, multi-processor system is used to designate which devices on a node should react to this request. On a Purge command if the field is set to a specific processor ID, all other on-node processors will purge their copies of the data and the designated processor will keep its copy. On all other requests, a destination field set to equal a processor would imply that only that processor should take action.

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

Purging Shared Cache Lines in a Multi-Processor Node

      The destination field of a request in our multi-node,
multi-processor system is used to designate which devices on a node
should react to this request.  On a Purge command if the field is set
to a specific processor ID, all other on-node processors will purge
their copies of the data and the designated processor will keep its
copy.  On all other requests, a destination field set to equal a
processor would imply that only that processor should take action.

      In our multi-node, multi-processor system each processor has
its own cache.  Memory is shared so data associated with a given
address may be in the caches of several different processors on the
same or different nodes.  Before a processor can get exclusive rights
to an address, all other cached copies must be purged.  The processor
initiates the action by sending an exclusive request to the coherence
unit associated with the address.  The coherence unit tracks all
nodes having copies of local addresses so it knows to whom it must
send purges.  The unit does not know which processors on a node have
the copies.

      Purges are then sent to all nodes with cached copies.  As with
all requests, the destination field is used to designate which
devices on the node should receive the request.  If all processors
are to act on the request as is the case with a purge command, the
field is set to broadcast.  The exception is the purge command that
returns to the originating node.  The source processor must know to
keep its copy while all other on-node processors must know to purge
theirs.  Therefore, the destination field cannot be set to broadcast
on this purge.

      To prevent this source processor from purging its copy, the
des...