Author: Robert Hyatt
Date: 13:42:13 07/19/00
Go up one level in this thread
On July 19, 2000 at 14:37:39, Mogens Larsen wrote: >On July 19, 2000 at 14:02:00, Robert Hyatt wrote: > >>Not that I recall. First, I don't like single-cpu tests, for anything other >>than what I use them for, namely debugging. If you insist on single-cpu tests, >>then I still prefer native-mode ponder=on. Ponder=off actually is worse with >>some programs. IE older chessmaster programs would use 100% of the cpu even >>with ponder=off. Because of a keyboard input loop. So in a ponder=off match, >>it would get roughly 75% of the total cpu cycles. Ponder=on at least makes >>that 'equal' assuming the operating system is sane about process scheduling. > >There's quite a few if's involves. The chessmaster bug wouldn't be helped by >ponder=on on one machine AFAIK. Actually it does. Because then chessmaster is 100% compute bound and gets 1/2 of the cpu cycles, where crafty is also 100% cpu bound and gets 1/2 of the cycles. That is about the best way to do this, particularly with so many variables among all the different programs... > There's equal chances of screwing up cpu >allocation whether it's ponder=on or off. Since I'm quite frequently testing >programs under development without pondering implemented, I see no reason to >divert from the native-mode of single cpu machines and use ponder=on. I haven't seen a recent O/S screw up process scheduling if you have two compute- bound processes running concurrently. In general, each gets 50% which is what is supposed to happen. > >>Oddball conditions are ok, if you want to >>know how programs do under oddball conditions. But to extrapolate from that to >>normal conditions is a real stretch of the imagination... > >Noone is extrapolating anything, because data exists on both cases. Until now, >the data suggests that the difference in strength between programs with ponder >doesn't differ significantly with the results obtained using ponder=off. As I >mentioned in my previous post, this correlation may only exist in a small >spectrum, but who knows. > >Best wishes... >Mogens Some do extrapolate. IE program X beats program Y 60%-40%. Someone else runs the test and says "I get 50%-50%, why are my results different?" Then you find out that it was a one-cpu match vs a two-machine match. The same problem affects memory. One machine means you typically "short" the programs on hash. And in the case of Crafty, phash and egtb cache. Which further alters the outcome.
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.