Author: Robert Hyatt
Date: 07:45:03 04/22/01
Go up one level in this thread
On April 22, 2001 at 03:51:45, Uri Blass wrote: >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. > This happens about 15% of the time in today's programs. That leaves 85% of the time where you will get very little. >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. Depends on whether the program uses null-move or not, whether it is selective or not, etc. But mathematically you can prove that the first move will take about 1/2 of the total iteration time even without null-move, etc. This goes down a bit in simple endgames where hashing kicks in... > >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. I can guarantee you that searching only at the root in parallel will not get you one ply deeper. It might in 1 or 2 moves out of every 10. But the other 8 moves will be to normal depth. I did this in 1983 in fact, for a two-cpu Cray. When we went to the 4 cpu machine and ran no faster, I got busy to implement something better. > >searching 7 legal move by searching only them also means searching in most of >the cases one ply deeper if you guess well. How do you figure this? I already predict/ponder the right move way more than 50% of the time against GM players. In those cases multiple pondering won't save a thing. > > 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 Two is pretty poor for 8 cpus of course. And then there is the problem of the parallel algorithm, parallel hash probes, repetition detection, and on and on...
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.