Browse Prior Art Database

Mechanism to Reduce the "Unnecessary Sharing" of Update-Type Tightly Coupled Multi-Processor Cache

IP.com Disclosure Number: IPCOM000120891D
Original Publication Date: 1991-Jun-01
Included in the Prior Art Database: 2005-Apr-02
Document File: 3 page(s) / 108K

Publishing Venue

IBM

Related People

Kawachiya, K: AUTHOR [+3]

Abstract

Disclosed is a mechanism for reducing the "unnecessary sharing" of an update-type Tightly Coupled Multi-Processor (TCMP) cache. The key idea is to introduce the TIMEOUT mechanism to invalidate the unnecessarily- shared data.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 59% of the total text.

Mechanism to Reduce the "Unnecessary Sharing" of Update-Type Tightly
Coupled Multi-Processor Cache

      Disclosed is a mechanism for reducing the "unnecessary
sharing" of an update-type Tightly Coupled Multi-Processor (TCMP)
cache. The key idea is to introduce the TIMEOUT mechanism to
invalidate the unnecessarily- shared data.

      Fig. 1 shows a typical configuration of the TCMP system -1s.
Each processor is connected to the shared bus through the "snoop
cache."  Protocols used to maintain the consistency among these
caches can be divided into two types -- update-type and invalidate-
type.  When a processor writes to a data item shared with other
caches, the update- type protocol updates other processors' caches,
while the invalidate- type protocol invalidates them.

      Essentially, the update-type protocol is more effective for
data that is REALLY shared among processors.  However, if a data item
that need not be shared is shared among caches, this protocol
degrades the system performance by continuing the unnecessary
updates.  This "unnecessary sharing" is caused by a data item that
remains in the cache after the processor has finished using it.

      The disclosed mechanism reduces this "unnecessary sharing" by
using timeout.  If a processor does not access a shared data item in
the cache for a definite period, the data item will be invalidated
(rather than be updated) at the next "update request" from the shared
bus.

      Fig. 2 shows ...