Author: Daniel Shawul
Date: 06:57:24 11/23/04
Go up one level in this thread
On November 23, 2004 at 08:56:00, Anthony Cozzie wrote: >It is indeed YBW :) > I assumed it was not YBW because i think i read in a paper(taxonomy of parallel search algorithms) that it needs non-recursive search and crafty isn't. I have now got some info in CCC archieves( It is a kind of recursive YBW i guess). PVsplit by marsaland is the only paper i read and it seems pretty easy for crafty. I downloaded the DTS paper and tried to roughly grasp the concept.Yes it is indeed a difficult one. Thank god PVsplit is not that bad at duals! You must have had difficult time implementing all that. daniel >IMO YBW is the only way to get a good speedup. You want to split at ALL (fail >low) nodes, otherwise you are wasting your time. At CUT (fail high) nodes most >odern engines will get a fail high on the first move 90%+ of the time. >Therefore, if you have searched at least 1 move its very likely that the node >will be an ALL node. > >DTS tries to split at ALL nodes as well, but Bob used some more complicated >stuff to try to figure that out, that I wasn't sure how to adapt to nullmove (I >just use moves searched, and (hint) the transposition table). DTS is more of a >set of ideas on how to make processors work together, a framework if you will, >then a splitting selection criteria. As soon as I read the DTS paper, I knew I >had to implement it, because of the crazy things that can happen in DTS: > >1. CPU A starts searching a node. It adds CPU B. CPU A runs out of moves. >CPU B returns from the node (note that this precludes a recursive search). > >2. CPU A is searching a node at ply 10. It wants to add CPU B, but ply 10 is a >bad place to split, so it adds it at ply 3. > >3. DTS actually does score updates. So if CPU A starts searching with alpha= >0.45 and CPUB finds a new best move, A's alpha will be updated appropriately. >Zappa doesn't bother, it just aborts nodes if necessary. > >anthony
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.