Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: speedup of cray blitz as published in 1997

Author: Robert Hyatt

Date: 14:48:07 09/25/01

Go up one level in this thread


On September 25, 2001 at 14:26:09, Vincent Diepeveen wrote:

>On September 25, 2001 at 14:09:38, Robert Hyatt wrote:
>
>>On September 25, 2001 at 11:24:16, Vincent Diepeveen wrote:
>>
>>>On September 25, 2001 at 09:56:52, Robert Hyatt wrote:
>>>
>>>>On September 25, 2001 at 07:28:17, Vincent Diepeveen wrote:
>>>>
>>>>>On September 25, 2001 at 00:40:53, Robert Hyatt wrote:
>>>>>
>>>>>>On September 24, 2001 at 23:45:30, Dave Gomboc wrote:
>>>>>>
>>>>>>>On September 24, 2001 at 22:30:26, Vincent Diepeveen wrote:
>>>>>>>
>>>>>>>>Hello,
>>>>>>>>
>>>>>>>>Here written down speedups as claimed by a guy called R. Hyatt
>>>>>>>>by cray blitz for 24 different positoins as they occured in
>>>>>>>>a game:
>>>>>>>>
>>>>>>>>pos  speedup
>>>>>>>>1     2.0
>>>>>>>>2     2.0
>>>>>>>>3     2.0
>>>>>>>>4     2.0
>>>>>>>>5     2.0
>>>>>>>>6     2.0
>>>>>>>>7     1.9
>>>>>>>>8     2.0
>>>>>>>>9     2.0
>>>>>>>>10    2.0
>>>>>>>>11    2.0
>>>>>>>>12    1.9
>>>>>>>>13    1.9
>>>>>>>>14    2.0
>>>>>>>>15    2.0
>>>>>>>>16    1.9
>>>>>>>>17    1.7
>>>>>>>>18    1.8
>>>>>>>>19    2.0
>>>>>>>>20    2.0
>>>>>>>>21    2.0
>>>>>>>>22    1.9
>>>>>>>>23    2.0
>>>>>>>>24    2.0
>>>>>>>>avg   2.0
>>>>>>>>
>>>>>>>>So YOU, Robert Hyatt, claims in an OFFICIAL magazine,
>>>>>>>>called ICCA journal march 1997,
>>>>>>>>an AVERAGE speedup of 2.0 with cray blitz at 2 processors.
>>>>>>>>
>>>>>>>>Now i claim the same with DIEP if i'm not using dangerous
>>>>>>>>extensions (which btw are turned on by default).
>>>>>>>>
>>>>>>>>It appears you hadn't turned them on either (smart guy
>>>>>>>>to publish only speedups without dangerous extensions and only
>>>>>>>>tell in 2001 that you hadn't turned them on).
>>>>>>>
>>>>>>>2.0 isn't a problem, it's >2.0 that gets people up in arms.
>>>>>>>
>>>>>>>Dave
>>>>>>
>>>>>>
>>>>>>It isn't even 2.0...   I don't have the paper handy now but I seem to recall
>>>>>>that it was 1.9.  Certainly it isn't 2.0 just by inspecting the numbers he
>>>>>>gave...
>>>>>>
>>>>>>And the 4 8 and 16 processor tests were worse.  But comparing them to crafty is
>>>>>
>>>>>When i tested with 4 i also never got > 4.0 with normal diep versions,
>>>>>so very consequent with Cray Blitz, of course there are good reasons
>>>>>why my speedup tests i get 2.0 in practical game
>>>>>play at tournament level versus cray blitz 2.0
>>>>>in practical game play getting the same at 2 processors.
>>>>>
>>>>>Bob explained me quite clearly how he had done things in Cray Blitz,
>>>>>my entire algorithm is based upon that of course!
>>>>>
>>>>>I am sure that if cray blitz would run on 2 faster Cray processors,
>>>>>or simply run longer, that then all its findings will be similar to
>>>>>my findings!
>>>>
>>>>
>>>>
>>>>Why don't you look at that JICCA article.  Those _were_ long searches.
>>>>Since I originally ran on 16 processors, I stuck with those results and
>>>>backed up to 1, 2, 4 and 8 processors.  the 1 processor searches took hours
>>>>at times.  The 2 processor tests were _not_ just 3 minute searches.  They
>>>>were huge.
>>>
>>>R=1, futility, 200Mhz things and most likely a relatively bad
>>>tuned evaluation, when compared to todays standards.
>>
>>I have said this once.  I will say it again.  You can take crafty, and
>>run some tests with R=0, R=1, R=2 and R=2-3 (or any other R values you
>>care to try).  That will _not_ affect the SMP performance in the least.
>
>Please do so on the 24 game positions as used in your Article about
>Cray Blitz. Also for every depth apart. To also see possible relationships
>in search depth.
>
>Each position for 1 hour or so. If automatically scheduled that's within
>24 hours.

That won't be 24 hours.  The test has to be run for 1 cpu and for 4 cpus
(and probably for 2 for completeness).  I see no reason to worry with 1
hour searches since that can't be done in real games.  In Cray Blitz, I
took a real set of searches on 16 cpus and ran the same tests on 1,2,4 and
8.  The 1 cpu tests took forever for obvious reasons.




>
>>The selectivity of the tree has nothing to do with the efficiency of the
>>parallel search activity.
>>
>>I will be happy to post a few numbers if you want...
>>
>>Evaluation also has no effect on parallel search performance.  You can be
>>bad or wonderful there and have a good or bad parallel search for either.
>>
>>
>>
>>>
>>>>
>>>>
>>>>>
>>>>>>very hard because of the different approaches to almost everything from move
>>>>>>ordering to search algorithm...
>>>>>
>>>>>Crafty is recursive, only when you have your own program parallel you
>>>>>will understand what a big difference in parallel speedup this means
>>>>>for DIEP + Cray Blitz. I do not know whether Fritz is recursive. It's
>>>>>assembly, so perhaps Frans can somehow avoid recursive problems in assembly
>>>>>in a smart way.
>>>>
>>>>
>>>>
>>>>I have done it both ways.  The advantages to non-recursive are only programming
>>>>issues, _not_ efficiency issues. There is nothing I could do in CB that I can't
>>>>do in Crafty, if so inclined.  I chose to avoid the extra complications for the
>>>>moment, but it _was_ a choice, not a requirement.



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.