Author: Robert Hyatt
Date: 14:53:24 09/04/02
Go up one level in this thread
On September 04, 2002 at 16:47:50, Gian-Carlo Pascutto wrote: >On September 04, 2002 at 12:00:32, Robert Hyatt wrote: > >>Have I not said _many_ times that the parallel speedup is a very dynamic >>value that can change significantly on the same position run multiple times? >>have I not given you several examples of such? Do you not understand that you >>can run a test once and get 2.8, and run it again and get 3.1? If you don't, >>I can't help you at all. > >Your problem is that my testing included error analysis (*), and as far >as I remember (again, dont have the data on this machine), the result >was *not* compatible with a speedup of 3.1. So? Do we have to keep going back to the variability. Did you see the log file from my 3.1 run? I sent it to vincent. I thought it went to everybody in the discussion. your "not compatible" is _meaningless_. You take that speedup as an absolute number. I only +wish+ it was so. But it isn't. I posted a single position yesterday with a 10% variance... I have some that are far worse. Which means there is no _absolute_ speedup number, like it or not. A speedup number is just like a FIDE rating. It applies to a particular situation only... IE Eugene has posted 4 different tests producing 1.9 on Intel boxes, and 2 tests producing 1.4 on AMD. Even the hardware can change things... > >The speedup for the same experiment *cannot* be two different values that are >outside each others error margins at the same time. Want to bet? The problem is your "error margin" is wrong. Different positions produce different sorts of speedups... If you want to say I can't produce 3.1, fine. But I sent the log. Will it be 3.1 the next time? No idea. Maybe more, maybe less. I'll be happy to post some positions that show this kind of variance... > >Your speedup in my test conditions was 2.8 plus or minus something. >Not 3.1 plus or minus something. OK. So what... I supplied raw numbers. We ran the same test on two different machines, I got 3.0, you got 2.8. You say my 3.0 is outside the error range? So it is _impossible_ as Vincent would like to say? Even though the log clearly shows it happened? :) > >(*) Which is missing in your papers *everywhere*. In fact, it's exactly >this that caused Vincent and me to discover your time numbers were >questionale. It is not "missing". It is discussed often. But trying to define the 'error' is non-trivial... it is easier to define it for a single position and multiple runs, than for several positions, one run each. > >-- >GCP
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.