Author: Uri Blass
Date: 00:51:45 04/22/01
Go up one level in this thread
On April 21, 2001 at 22:52:42, Robert Hyatt wrote: >On April 21, 2001 at 11:20:46, Uri Blass wrote: > >>On April 21, 2001 at 09:02:29, Ed Schröder wrote: >> >>>I have promised to give an update about my talks with Enrique the organizer of >>>the computer - Kramnik event. Enrique said no. His reasons are the contractual >>>obligations to have the play-off started on April 26 (next thursday) meaning no >>>delays. >>> >>>On the question to have the programs ready before April 26 Enrique questioned >>>the strength of the multi versions of Chess Tiger and Rebel. >>> >>>As a result I gave up. >> >>I think that it is not fair from Enrique to do it. >> >>I guess that the multi versions of Rebel or Tiger earns less from more >>processors but still earn something from them because it is impossible to >>optimize programs for more than one processor in a short time. > >We could call this "A requiem for disaster". A parallel search is _not_ >something you throw together in a few weeks or months. Not and take it to >the most visible event in years. I've done my fair share of parallel search >chess engines. None were easy. All had bugs that took months (or even years) >to find. Jumping on to the bandwagon in a quick and dirty way is not the way >to go and if I were doing such a "qualification event" with limited time and >resources I wouldn't accept such a program either. I don't particularly believe >in the "qualification event" as described anyway, but a brand new parallel >search program is not particularly exciting to think about either... > >Even with a parallel search program that has public source, copying the >search is non-trivial. Understanding it is something else. And merging >it with a significantly different engine would not be easy. > >So copying would be hard. Development from scratch would be hard. Sounds >like a mess... > > > > >> >>Am I right? >> >> >>Here is an idea how to use 8 processors in a simple way. >> >>Give one processor to analyze in the regular way(I will call it processor A). >>Guess 7 candidate moves to be the best move and give the other 7 processors to >>analyze only the candidate moves(one move per processor). >> >>If proccesor A do not suggest one of the 7 candidates move as best and if the >>score of the move of it is better than the scores of the other processors then >>play the move of processor A >> >>In the other cases play the move with the best score based on the scores of the >>7 processors. >> >>If you can guess correctly in most of the cases then it means that you can >>search 1 ply deeper in most of the cases thanks to the 8 proccesors. >> >>Uri > > >You won't get one ply deeper. The first move searched at any ply takes >about 75% of the time in normal cases. The event when a program changes it's mind in the last ply is not rare and in these cases the first move usually does not take 75% of the time. My impression is that even when the program does not change it's mind the first move searched takes less then 75% of the time if you wait to finish the iteration. searching all the legal moves one ply deeper means usually getting one ply deeper except cases that the right move is not a quiet move. It may mean more then one ply if the right move is pruned by null move pruning. searching 7 legal move by searching only them also means searching in most of the cases one ply deeper if you guess well. Such a parallel search would maybe >get a speedup of 1.5 if you are lucky. I am more optimistic and I believe it is possible to get speed up of about 2 by this way if you use 8 processors. You can get good guesses for 7 processors based on previous search. Uri
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.