Author: Uri Blass
Date: 04:47:40 10/04/01
Go up one level in this thread
On October 04, 2001 at 07:09:30, Uri Blass wrote: >On October 04, 2001 at 06:46:48, Sune Fischer wrote: > >>On October 04, 2001 at 04:56:08, Uri Blass wrote: >> >>>The difference is that the other side(I and Bruce) do not say that they know >>>that there is a super linear improvement but that we cannot say that it is >>>impossible with the known algorithm of today. >>> >>>I said that the only way is to investigate the problem by test positions in >>>order to see if programs can get a super linear improvement from 2 processors. >> >>But that is not the only way, you can also use logic. >>This is like explaning to an inventor why he can't make a perpetuum mobile >>machine. If he doesn't understand the laws of physics, does not know of energy >>conservation, then he will keep arguing till the day he die, that "we can not >>know for certain until we have tried everything". >> >>We _do_ know, there is proof and Bob has outlined it several times, but if you >>won't listen or understand, then we have a communication problem. >> >>-S. > >There is no proof when we do not know the source code of Deep fritz. > >There is an agreement that a poor sequential algorithm can be more than 2 times >slower than a good parallel algorithm with 2 procesors. > >We do not know that the algorithm of Deep Fritz for one processor is not poor > >It is good enough to lead the ssdf list but it does not prove that it cannot be >improved and it is possible that the programmers of Deep Fritz even do not know >that it is possible to improve it at long time control because they did not test >it for that purpose because they were interested at tournament time control and >not at slower time control. > >In the positions that I posted Deep Fritz needs a long time to find the >suggested move(in some cases many hours) and always more than 20 minutes on p800 >if I remember correctly. > >Uri Another note about it is that I admit that I do not understand the way that parallel search works but it is not important to understand that impossible is not something that I can say for source code that I do not know because I am not sure if Deep fritz is using the same algorithm as Crafty. Bob can at most say that it is impossible for crafty but I understand that even for Crafty there are positions when parallel search can give programs super linear improvement. The question is what is the reason for the super linear improvement of Crafty in some positions and how it is possible to prove that there cannot be average superlinear improvement for the existing Crafty at long time control that Crafty was not tested(I agree that in this case it is possible to improve Crafty for one processor at long time control but the discussion is about existing programs and not about science that talk about theoretic programs). Another note is that an average super linear improvement does not mean always doing better than linear improvement. If I have a way by 2 processors to be 4 times faster in 1/2 of the cases and 4 times slower in 1/2 of the cases then I have a super linear improvement but I think that it is still better to be twice faster in all the cases. 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.