Author: Robert Hyatt
Date: 08:38:33 12/23/03
Go up one level in this thread
On December 23, 2003 at 10:08:53, Roberto Nerici wrote: >On December 23, 2003 at 09:03:45, Anthony Cozzie wrote: > >>On December 23, 2003 at 06:39:22, Peter Berger wrote: >> >>>On December 22, 2003 at 22:59:01, Robert Hyatt wrote: >>> >>>>>To go back to the discussion that started this: it looks as if Crafty on 10 >>>>>times faster hardware is indeed very competitive with a top commercial engine, >>>>>but not necessarily the overwhelming favourite. >>>> >>>>That was my original point in response to Omid's rather crass "if they >>>>thought they had any chance of winning, they would have come..." statement. >>>> >>>>And the main point is that I probably would not have been "just" 10X >>>>faster. :) >>>> >>> >>>Interesting. I think this was not challenged in discussion partly because there >>>was no clear idea how much faster Crafty would have to be. >>> >>>Let's say you had planned to show up with 10X faster hardware with Crafty at >>>WCCC2003. >>> >>>This is what Junior used there: >>>http://www.chess.at/turniere/turniere2003/chess003/video/int2.wmv >>> >>>Intel 4* 2.8 GHz >>> >>>What would you have got for Crafty? >>> >>>Peter >> >>It would probably not be too hard to get a 64-way Itanium2 box. I know Vincent >>was "practicing" on one of those :) I think that'd be about 60M nps for crafty. >> I'm not sure how much stronger it would be, though. I think evaluation is much >>more important than search depth once you start hitting 14-15 ply. >> >>anthony > >I doubt it. Bob has had to put a lot of work into getting really good scaling >for a 4-way Opteron (he's been very successful, however). Last time he posted >about this, he was getting nearly 7Mnps on Windows and ~6Mnps on Linux. > >I don't know what you would get if you just stuck the current version on a >64-way machine, but I doubt it would be anything near 60M. Actually part of the NUMA changes and testing Eugene did were on an Itanium box. It is not a bad machine, for the record... > >(Actually, someone did post to the Crafty mailing list some results they'd got >from running an earlier 19.x version of Crafty on some _really_ big machines. I >don't have the post to quote from, but they ran the benchmark on a different >number of processors and found the total nps was less on 8 processors than on a >lower number.) Correct. That is what happens when you run on a NUMA box. The problem with NUMA is that it is also non-standard in how it is supported by each vendor, which is in stark contrast to something like POSIX threads or whatever, which has a well-defined API. NUMA has a well-defined set of needed capabilities, but the API to get to those capabilities is a hodge-podge of hacked in things that make portability tough. All I can say is that I had already studied the NUMA-related issues within Crafty when I was working on a NUMA-alpha machine with COMPAQ, and it worked pretty well on a 64-way alpha box. No doubt some tweaks would have been needed to run on a non-windows platform, but Eugene's NUMA code and the Crafty changes I made to use it should work pretty decently on any windows NUMA box, which includes both Opteron and Itanium. The alpha NUMA stuff is completely different since they don't use windows. > >So could Crafty ever get 10 times the nps than Junior on a quad Pentium? I can't >see how, as things currently stand. Crafty was doing 10M+ (endgame) on a quad NUMA opteron. The scaling (NPS increase per processor) was _very_ high, with 4 producing over 3.9X the NPS of one. The test you mentioned earlier was way before either of us had tried to test and tune for a NUMA box, other than the alpha using unix... > >Roberto/.
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.