Author: Mogens Larsen
Date: 03:50:23 09/20/00
Go up one level in this thread
On September 19, 2000 at 11:50:02, Robert Hyatt wrote: >Because move ordering is not perfect. If you start a parallel search at a >ply where only one move needs to be searched, the extra processor is searching >a part of the tree that a single processor program would not search due to an >alpha/beta cutoff. Since move ordering is not perfect, it is impossible to >accurately choose a position where all moves have to be searched, vs a position >where only one move needs to be searched. > >As I have mentioned before, in Crafty, I am about 92% right. That is, if I am >going to get a beta cutoff, 92% of the time it happens on the _first_ move >searched at a ply. Unfortunately, that 8% where it is wrong is a real killer >to parallel performance. > >Hence my 30% per processor overhead estimate. 1cpu = 1.0 speedup. 2cpu = 1.7 >speedup, 3cpu = 2.4 speedup and 4 cpu = 3.1 speedup. All rough averages, actual >numbers will vary wildly. Isn't it possible to make the processors work independently of each other? Eg. one making a knowledgebased positional search, while the other one searches for tactical possibilities, maybe with more selectivity. Examining primarily the move searched by the positional engine for tactical difficulties. Like Little Goliath and Hiarcs rolled into one :o). Would that be possible at all given that an evaluation function would be able to interpret the two search results and make a correct selection. Mogens.
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.