Author: Vincent Diepeveen
Date: 10:59:40 06/22/98
Go up one level in this thread
On June 22, 1998 at 11:26:42, Robert Hyatt wrote: >On June 22, 1998 at 10:50:01, Don Dailey wrote: > >>On June 22, 1998 at 09:51:27, Robert Hyatt wrote: >> >>>On June 22, 1998 at 02:37:40, Roland Pfister wrote: >>> >>>>I was on saturday at the townhall to watch Anand - Fritz and perhaps >>>>meet some computer chess freaks. >>>> >>>>But the only persons I knew were the two Matthiasses from ChessBase >>>>and Frederic Friedel. Frederic was busy photgraphing or videoing, so >>>>I had a small talk with Matthias W. and he told me that Fritz played the >>>>open too, at that moment having 5 points out of 5. Not bad. >>>> >>>>Then Mathias F. arrived, he was the one to operate Fritz. I asked him >>>>how they did parallize Fritz and he answered "MTDF", because they >>>>did not have the time to do something clever like Bob. >>> >>>this sounds like they did something very ugly, like searching a different >>>mtd(f) on each processor? But in any case, if they used mtd(f), then we >>>didn't see the "real fritz" play chess here, because mtd(f) is also not >>>quick and dirty to implement. It requires lots of algorithm changes, it >>>*definitely* requires significant hash-table changes, and so forth. Seems >>>like they were too interested in going faster, but overlooked difficulties >>>in getting stronger. >> >>I think mtd(f) is a pretty logical choice for a parallel machine >>if you were in a big hurry. I think the key thing that they must >>have considered was how clean and simple the algorithm is and >> > >However, remember that I played with this too, and there are plenty of >issues to keep someone busy for several weeks to get it right. And if it >is not "right" it will be slower than one processor using normal PVS. > >the zero width window searches which is ideal for parallel chess. >> >>Yes, there are side affects that eventually create lots of >>implementation work later but these are not so serious that >>cannot be ignored (if you are in a big hurry.) >> >>I'm guessing that they are not doing what you conjectured, but are >>simply spawning off sibling searches in parallel with of course >>mtd's zero width window. I wonder if they have abort in the >>code or simply wait? I can imagine these considerations might >>very well push you to consider mtd(f) in a parallel program. >> > >that's what I conjectured, in fact, that they are doing the mtd(f) >iterations at the root in parallel with each processor searching a >different "window". Doesn't sound particularly good, however, based on >normal performance considerations. > > > >>Another possibility for easy implementation is mtd(f) with >>young brothers wait, then spawn off siblings in parallel with >>no aborting. Not having to worry about re-searching must also >>be a desirable characteristic for a quick and dirty >>implementation and another reason they might have considered >>mtd(f). >> > > >If they had done this, they could have used the code directly from Crafty >as it quite easily does this, although the data structures are on the >complex side. But the statement "we chose mtd(f) rather than something >more clever like Bob did" seems to say they don't split *inside* the tree >at all... > >>- Don Note that one of the fritz team namely Frans Morsch has to make MTD in assembler then, and i wouldn't know a single reason why Frans would make it. Parallel programs aren't that demanded. I doubt whether Chessbase runs the search parallel. I think it's more like some within chessbase realized that different engines can run in parallel, and that made directly a good publicity stunt for them and Siemens, further there is a small chance that they do something with the first move. Meanwhile i'm busy to implement parallelism into Diep, i'm using DTS as described by Hyatt, with this difference that i'm trying to prevent to copy the whole search tree (as this ain't cheap at intel PC).. 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.