On 7 Feb 2003, Perry E. Metzger wrote:

:
:An interesting article on how to improve spin loops on intel
:hyperthreading processors.
:
:http://cedar.intel.com/cgi-bin/ids.dll/content/content.jsp?cntKey=Generic+Editorial%3a%3amp_spinloops&cntType=IDS_EDITORIAL&catCode=BZC
:
:Introduction
:
:Parallel programs with multiple threads must use synchronization
:techniques in order to insure correct operation. Generally,
:synchronization operations use shared synchronization variables and
:"spin-wait" loops that check on the values of those
:variables. Starting from the Intel? Pentium? 4 and Xeon? processors,
:Intel? IA-32 architecture provides a new instruction to address the
:performance issues associated with spin loops.
:
:This application note addresses two important optimization issues for
:multi-threading computations involving high-speed processors: spin
:loop and shared-data management. Specifically, these optimizations
:include the use of the new PAUSE instruction in spin-wait loops and
:the placement of shared and non-shared data on different 128-byte
:cache lines. Intel strongly recommends using the new PAUSE instruction
:in spin-wait loops as soon as possible since it is backward compatible
:with all earlier IA-32 architecture. This document describes in detail
:the recommended changes and the reasons behind these changes.
:

FWIW, the FreeBSD locking primitives do use PAUSE if you're interested in
how/when they do it.  Just a reference :)

Cheers,
Andrew

--
Andrew R. Reiter
arr@watson.org
arr@FreeBSD.org