Author: Robert Hyatt
Date: 14:27:00 09/19/03
Go up one level in this thread
On September 19, 2003 at 04:15:49, Gian-Carlo Pascutto wrote: >On September 18, 2003 at 19:06:54, Eugene Nalimov wrote: > >>>>>>No there isn't, because for single and dual machines, the NUMA issue can >>>>>>pretty well be ignored >>>>> >>>>>Ignored? You _will_ get a performance handicap if you do so. >>>>> >>>>>-- >>>>>GCP >>>> >>>>Yes, but on a dual it is a _minimal_ handicap. >>> >>>Last week, this 'minimal' handicap was enough to require a special >>>version of Crafty. Now it's not needed for the Linux kernel? Uhh... >> >>I believe you never participated in a *huge* software project. >[irrelevant blabla] >>And so on, and so on... > >This is all fine and dandy, but it has absolutely nothing to do with the >discussion at hand. > >Bob has been going on for a week about how he was going to make Crafty NUMA >because that last 10% (or less) was so important, that it certainly _couldn't_ >be ignored. "shouldn't be ignored". Also I haven't "been going on for a week about how I was going to do _anything_." The question was asked. I answered it. > >Now he's saying the exact same thing is irrelevant, and the Linux kernel >people can ignore it. Where did I say that? I _never_ said it. I said that linux on a dual NUMA box works pretty well "as is". On a larger box it needs work. pretty simple statement, to me... > >I realize all too well the Linux kernel is a bigger project, but it's also >got quite a few more people working on it, quite a few more testers, and >quite a few more users than Crafty NUMA will ever have. > >For Crafty, that last 10% is so important, but for a project that's used by >1000x of times more users, it's irrevelant? That just makes no sense! Only because you aren't following the conversation... nobody said what you are claiming. > >Whether or not it takes a lot of work is completely irrevelant. This discussion >is about whether you can ignore the small speed advantage or not. Of course >it'll take a while to implement and bring problems! But do you _have_ to do >the effort or not? On a dual NUMA box, the O/S issues are far less worrisome than for a 32-cpu box (or larger). I'm not going to take the time to go into why unless it is really necessary. For crafty, it will run on a dual NUMA box with no problems. Eugene can give you some numbers for a NUMA Itanium if nis NDA is not still in force. He can also tell you what happened when he tried more processors... > >The real world (in case the Linux people) seem to disagree with you as well, >because they're quite well aware of their problems with NUMA architectures >as simple as the Opterons and _are_ working to improve the support. > >There will be a lot of Linux servers running on Opterons and having a >performance handicap as compared to Windows boxen in that regard is not >something that is acceptable for them. > >-- >GCP Of course it isn't, but at the moment there are not a "log of opterons" around. The demand is low, for the moment.
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.