Author: Robert Hyatt
Date: 07:52:40 08/06/05
Go up one level in this thread
On August 06, 2005 at 04:31:50, Gian-Carlo Pascutto wrote: >On August 05, 2005 at 12:48:39, Robert Hyatt wrote: > >>Personally I am giving up on Red Hat. The enterprise edition we were using had >>too many bugs. CPU time was hopelessly wrong (8 threads was reporting 7x the >>cpu time that was possible at times, etc. Suse works... > >One difference of note is that Red Hat is using NTPL threads and SuSE is using >the LinuxThreads system. > >NTPL is (much!) higher performance but also newer and probably more buggy, too. > >-- >GCP I have not run on both. No difference. Works perfectly on non-dual-core processors. There are some subtle cache differences. Now there are two L2 cache controllers attached to a multiplexor to the hypertransport. On a single-core there is one L2 per hypertransport. Whether that is the issue that is smacking me or not is currently unknown. I am currently re-doing my older "non-threads" code where I didn't use the threads library at all, I "rolled my own" using fork/clone... It is possible that the "control thread" that posix threads uses is causing some cache coherency traffic issues. Hopefully I'll have this new version working by tomorrow and will know more. Note that this does not seem (to me) to be a hardware "bug". It is an "issue" that is exacerbated by the second core, somehow. Whether it is something in Crafty, or outside Crafty, is presently unknown. I should have never removed the no-threads code I wrote the last time posix threads was broken by redhat. Unfortunately I did... life goes on...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.