Author: Robert Hyatt
Date: 08:53:32 05/07/04
Go up one level in this thread
On May 07, 2004 at 07:03:55, Mikael Bäckman wrote: >On May 07, 2004 at 04:30:09, Vincent Diepeveen wrote: > >>On May 06, 2004 at 12:45:43, Robert Hyatt wrote: >> >>Bob let's get realistic. >> >>BK is a flawed testset to test parallel speedups at. > >Why? I know nothing of parallel processing, but what in the BK makes it bad for >parallel testing? > >/Mikael > Vincent's declaration. Nothing else... > > >> >>What happened is. You took 4 positions from that to 'proof' your 3.1. >> >>That gets disproven then by GCP doing statistical math where you know really >>less than a waterbird from. You do not even realize what the +- behind every >>measured speedup means. In fact you never provide them yourself. >> >>GCP then is doing different tests at your quads and very clearly determines 2.8 >>using a-symmetric king safety. >> >>My 1.0 out of 2 speedup for crafty came of course from using tests with >>symmetric king safety. >> >>But GCP using 30 positions gets down to 2.8 speedup. >> >>And he gets one time 3.2 and another time 3.0 for speedup when searching >>fullwidth. >> >>a) this shows how poor your way of representing things is. You are just showing >>things like you want them to see >> >>b) i do not believe a crap you did a honest test where at position 21 in BK the >>crafty on average gets a 2.4 speedup when searching fullwidth. No matter how i >>test fullwidth at position 21, speedups are better there for fullwidth using >>asymmetric king safety. >> >>You did not do a honest testing to get to your numbers, even in those 4 >>positions and yet you deny GCP's testing at 30 positions which were honestly >>done, to be true. >> >>This shows your true nature. >> >> >> >>>On May 06, 2004 at 12:19:13, Gian-Carlo Pascutto wrote: >>> >>>>I hope you don't mean the ones blow. >>>> >>>>Are you still claiming you 'measured' 3.1 which supposedly contradicts >>>>the 2.8 I measured? >>> >>>I don't know what any of that refers to. My formula came from running the BK >>>test (23 positions, excluding #1 an instant mate in 3). >>> >>>I am trying to get the disk set up and installed now that I sent to AMD for the >>>last CCT event. While I had access, I ran the BK test to get the data. I had >>>promised Martin that I would post the numbers. >>> >>>But you _really_ need a better vocabulary. I have not tried to "contradict >>>2.8". I have _clearly_ said that the speedup varies and that 3.1 is the value >>>suggested by a linear approximation fit to a non-linear function. >>> >>>I do _not_ understand the obsession with "is it 2.8 or 3.1"? It could well be >>>_both_. >>> >>>I have _always_ called this an approximation. Let me get the log files and grep >>>the times for Martin. I'll put the logs on my ftp machine since they will be >>>fairly long, if you want to see the opteron data for 1-4. >>> >>>> >>>>>Return-Path: <gcp@sjeng.org> >>>>>X-XS4ALL-To: <diep@maildrop.xs4all.nl> >>>>>Date: Fri, 9 Aug 2002 09:14:59 +0200 (MEST) >>>>>From: Gian-Carlo Pascutto <gcp@sjeng.org> >>>>>X-Sender: <giancarlo@garf.natrese.be> >>>>>To: "Robert M. Hyatt" <hyatt@cis.uab.edu> >>>>>Cc: <diep@xs4all.nl> >>>>>Subject: Re: Parallel results so far >>>>> >>>>> >>>>> >>>>>On Thu, 8 Aug 2002, Robert M. Hyatt wrote: >>>>> >>>>>> Here is the results: >>>>>> >>>>>> ----------------------------------------------- >>>>>> ---null=2/3--- ---null=off--- >>>>>> position 1cpu 4cpu S/U 1cpu 4cpu S/U >>>>>> kopec 21 27.9 10.7 2.6 30.3 12.4 2.4 >>>>>> kopec 22 22.5 6.1 3.8 26.0 7.5 3.5 >>>>>> kopec 23 33.5 11.2 3.0 20.9 6.4 3.3 >>>>>> kopec 24 18.1 6.0 3.0 26.2 8.3 3.1 >>>>>> >>>>>> ----------------------------------------------- >>>>>> note. all positions were searched for 30-45 seconds >>>>>> with the last 1-cpu output used to measure how long >>>>>> the 4-cpu search took to reach the same output (say >>>>>> the end of a search, or a PV move and score displayed). >>>>>> >>>>>> Vincent claimed "I never ran this test." Thought I would >>>>>> run it _again_ just to expose "baloney". >>>>>> >>>>>> I think the conclusion from the above is >>>>> >>>>>Conlusions from the above? Howso? >>>>> >>>>> speedup >>>>>Nullmove 3.1 +- 0.25 >>>>>Non-nullmove 3.1 +- 0.25 >>>>> >>>>>The standard errors (1SD) are way too huge to allow what >>>>>you try to conclude. I measured a speedup of 2.85 with >>>>>nullmove and 3.1 without, whereas your test wouldn't even >>>>>be able to differentiate between the two. >>>>> >>>>>If you want to scientifically settle this, >>>>>you'll need more and better data. >>>>> >>>>>(I couldn't find the CCC article reffered to earlier) >>>>> >>>>>-- >>>>>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.