Browse Prior Art Database

Broadcast of Mostly-Read-Only Highly Shared Cache Lines in Multiproces Sor Systems

IP.com Disclosure Number: IPCOM000039210D
Original Publication Date: 1987-May-01
Included in the Prior Art Database: 2005-Feb-01
Document File: 1 page(s) / 12K

Publishing Venue

IBM

Related People

Robinson, JT: AUTHOR

Abstract

This article describes a method for identifying mostly-read-only highly shared cache lines in multiprocessor systems and for avoiding misses by most processors. Consider a multiprocessor system with N local caches, one per processor or local group of processors (there may be other levels of caches). For large local caches, it is expected that there may be some number of highly shared mostly-read-only lines that are referenced so often by each processor that they would remain permanently resident in all local caches were it not for occasional writes to these lines. In existing and proposed systems, when a write to such a line does occur on some processors, all other copies of the line are invalidated, resulting in a miss on the next reference to the line for each of the other local caches.

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

Page 1 of 1

Broadcast of Mostly-Read-Only Highly Shared Cache Lines in Multiproces Sor Systems

This article describes a method for identifying mostly-read-only highly shared cache lines in multiprocessor systems and for avoiding misses by most processors. Consider a multiprocessor system with N local caches, one per processor or local group of processors (there may be other levels of caches). For large local caches, it is expected that there may be some number of highly shared mostly-read-only lines that are referenced so often by each processor that they would remain permanently resident in all local caches were it not for occasional writes to these lines. In existing and proposed systems, when a write to such a line does occur on some processors, all other copies of the line are invalidated, resulting in a miss on the next reference to the line for each of the other local caches. First, consider the problem of identifying such lines when a write occurs. A storage director maintains the status of each line in each local cache, and is responsible for invalidating all copies of the line. Suppose that when the write occurs, the line was previously held read-only in all local caches (this can be determined simply by ANDing together a set of bits). Then for N >> 2, simple probability theory indicates that this line is almost certainly a mostly- read-only highly shared line (with increasing certainty as N increases). This criterion can be generalized by requiring that some number...