Author: Robert Hyatt
Date: 22:06:21 05/07/04
Go up one level in this thread
On May 07, 2004 at 13:24:10, Robert Hyatt wrote: >On May 07, 2004 at 12:13:45, Gian-Carlo Pascutto wrote: > >>On May 07, 2004 at 11:53:32, Robert Hyatt wrote: >> >>>Why? I know nothing of parallel processing, but what in the BK makes it bad for >>>>parallel testing? >>>> >>>>/Mikael >>>> >>> >>>Vincent's declaration. Nothing else... >> >>A certain article from a certain Dr. Hyatt explains the three classes of >>positions as far as parallel efficiency are concerned. I'm pretty sure BK >>(for a modern program on modern hardware) almost entirely falls into the >>'good' category :) >> >>-- >>GCP > >I have data for BK as well. Want to make a wager before you see it? I _know_ >what happens in the BK tests as I have run the positions _many_ times. There >are _plenty_ of positions (most in fact) where the PV best move changes, or >where the PV itself changes drastically, either of which hurts performance. > >Of course, if you have data where the BK tests produces a speedup far better >than the speedup on the DTS positions, feel free to share it. > >This is not the place for _speculation_ when actual factual data can be >presented with much more impact... > >Remember that 11 of the positions are tactical, 12 are pawn lever (I never test >#1 as it is a 3-ply position). There's +plenty+ of room for move ordering >problems. Even my favorite position, BK22, Bxe4, does not produce a very good >speedup as the Bxe4 move is found quickly, but the PV keeps changing and the >score keeps dropping, iteration by iteration. It isn't "stable" as I would >label "good" positions.. See the new logs on my ftp machine and the summary I posted a few mins ago. The BK positions are _not_ as good for parallel search as our resident expert about nothing Vincent has always claimed. The CB positions from the DTS paper are actually better for Crafty's parallel search performance by a small amount...
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.