Author: Vincent Diepeveen
Date: 19:21:21 05/08/02
Go up one level in this thread
On May 08, 2002 at 20:30:16, Nolan Denson wrote: >I was wondering which parallel search method is best. Some earlier posted the >info for me about which crafty was the first to use SMP ... (that version was >15.0 ) Now after looking into it the note states it uses the "PVS" algorithm >(Principal Variation Search). Now i know that Amy uses ABDADA as a parallel >search algorithm. So my ??? is > >1. Which algorithm is Crafty using today ??? > >2. Which algorithm is considered best??? > >3. What are some other algorithms thats are being used by some other dual > programs??? > >Anyone feel free to answer. > >Thanks > >Nolan Denson the best algorithm by far is what i'm using. but could be a year of work to get implemented correctly. took me months to get right. 1.5 years later i fixed another rare race bug (though it already hardly crashed). i'm using a pc modified version of Cray Blitz + YBW + new ideas. But don't get excited about the 'new ideas' section. it was pretty trivial. The non trivial thing is to implement bob's Cray Blitz algorithm. Bob can explain how it works. note he already did a bit in an article in ICCA, but he probably is happy to explain again here the basic concepts of parallel search. Without talking for months with Bob, and all kind of help (as i didn't have a dual machine even, because one of my cpu's was needed by office which was owner of both cpu's actually, and they needed one of them in a machine, and i didn't have money to replace it) i would never have ever managed to get my thing parallel to work. So much small details do not get mentionned in the 'papers'. there is no practical notes on it. I was writing an article for one of the coming icga with a practical notes on parallellizing a search using the Huber concept (simplified and better working lambdada or abdada search, i can't spell or remember that french name ever without looking into the icca journal...). Before you start implementing a concept, please consider this: a few years ago the Huber idea wasn't very well known yet. Why not start getting a simple speedup? If you make a chessprogram you don't start making a parallel searching thing with difficult mobility concepts in evaluation and difficult king safety terms directly in eval either. You start making a move generator in general. Best regards, Vincent
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.