Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Crafty multiprocessing update

Author: Robert Hyatt

Date: 06:51:37 06/09/98

Go up one level in this thread


On June 09, 1998 at 08:48:28, Rickard Björklund wrote:

>On June 08, 1998 at 23:17:25, Robert Hyatt wrote:
>
>>Young Brother's Wait, if you haven't seen it, simply says to not search
>>any moves at a ply (in parallel) until at least one move has been
>>searched
>>sequentially.  The obvious idea is to avoid search overhead where we
>>might
>>search moves before we have a good alpha/beta bound.  I've been playing
>>with
>>this and have found that a slight modification is currently working even
>>better for me, namely to not search in parallel until *two* branches
>>have
>>been searched sequentially, which improves the odds that if two moves
>>have
>>been searched, then *all* moves will need to be searched.  I have not
>>yet
>>tried 3, 4, ..., but am going to do so as this is easy and trivial to
>>change.
>>
>>More as new results and tests are available...
>
>Can a new hash table flag be useful here ? To tell if it's "safe" to
>search all moves in parallel at once, or if one should first try one
>or two moves in parallel ?
>
>Rickard

not easily, because this flip-flops from iteration to iteration. and
depending on whether the position occurs below a PV node or a non-PV
node.  Too many variables to predict... it is easier to hide behind the
door and watch a bit, then make a decision.  :)



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.