Author: Robert Hyatt
Date: 07:07:19 07/09/04
Go up one level in this thread
On July 09, 2004 at 09:47:36, Tord Romstad wrote: >On July 09, 2004 at 09:13:31, Robert Hyatt wrote: > >>On July 09, 2004 at 08:38:31, Tord Romstad wrote: >> >>>I am currently writing a chess engine. Parallel search is not among my main >>>interests at the moment, but it is not entirely impossible that I will give it >>>a try some time in the future. >>> >>>In order to keep everything as flexible as possible, I would like to design >>>my algorithms and data structures in such a way that adding parallel search >>>at a later stage is feasible. I understand that I should remove most of my >>>global variables and replace them with huge structs containing the same data, >>>and use one such struct for each thread. Is there anything else which is >>>important to keep in mind? >>> >>>Tord >> >> >>That's the main issue assuming you are going to use lightweight processes >>(threads) which I believe is the best approach. > >Yes, threads is what I intend to use. > >>The most thread-specific data >>you have, which means less global data, will help performance (modified global >>data is not cache-friendly on a SMP box) and simplify testing (since modified >>global data requires atomic locks to avoid interleaved update problems). > >But as far as I can see, I would be forced to pass an extra pointer to most >of the chess-related functions (search(), evaluate(), generate_moves(), >make_move() etc.). A bit ugly, but I suppose I could live with that. > >Thanks for your reply! > >Tord I don't see a better solution. It cost me about 3-4% slowdown when I first did this. It is probably significantly less than that on the opteron with 8 extra registers. The advantages offset some or all of that, as well. IE non-duplication of data that is computed at startup. Separate processes will have multiple copies of such data, such as egtb cache, etc... With threads there are no multiple copies of _anything_...
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.