Author: Sune Fischer
Date: 03:37:26 09/14/03
Go up one level in this thread
On September 13, 2003 at 20:09:09, Jeremiah Penery wrote: >If you want to calculate the ratio of effective nodes/time, I'm not really sure >how you plan to do it. How do you define an effective node, first of all? The >serial search duplicates nodes often enough on its own that I don't see a way to >calculate a difference between wasted serial nodes and wasted parallel nodes. >Or am I just completely off-base here? Wasted parallel nodes is wasted serial nodes *plus* parallel search overhead. We can't do a lot about the wasted serial nodes, because we'll never have a 100% perfect move ordering and pruning strategy, so let's not worry about how to measure that (right now) :) What's important is that we can look at the nps in a parallel search and directly compare them to the nps in a serial search. Take hyperthreading, Crafty apparently gets 30% higher nps so people who doesn't know about parallel search overhead might think this means it will solve things 30% faster. I would like this number corrected so that it showed how much faster we should expect it to solve positions, probably only 10-15 percent faster, so the nps should be 115% higher, and not 130%. Afterall it's not about getting high nps, it's about getting maximum efficiency. I'm not sure how hard it would be to implement though... -S.
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.