Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Let's talk about fraud.

Author: Robert Hyatt

Date: 08:58:21 05/03/04

Go up one level in this thread


On May 03, 2004 at 11:05:25, Vincent Diepeveen wrote:

>On May 03, 2004 at 09:20:51, martin fierz wrote:
>
>>On May 03, 2004 at 02:14:53, Vincent Diepeveen wrote:
>>
>>>On May 02, 2004 at 18:49:38, Robert Hyatt wrote:
>>>
>>>>On May 02, 2004 at 18:23:44, Gian-Carlo Pascutto wrote:
>>>>
>>>>>On May 02, 2004 at 13:12:04, Robert Hyatt wrote:
>>>>>
>>>>>>He sent me an email trying to justify his poor performance.  He first claimed
>>>>>>that it was an artifact of null-move.  Testing disproved that.
>>>>>
>>>>>What testing?
>>>>>
>>>>>--
>>>>>GCP
>>>>
>>>>
>>>>The testing you and I both did.  It showed a minimal speedup difference if you
>>>>recall.  2.8 vs 3.1...  not _that_ significant...
>>>
>>>2.8 for with nullmove
>>>3.0 for without nullmove
>>>
>>>A major difference. based upon 30+ positions.
>>>
>>>And both not *close* to speedup(n) = 1.0 + 0.7(n-1)
>>
>>i know nothing about this thread, i know nothing about multiprocessing, but i do
>>know that the above formula gives 3.1 for n=4.
>>i don't know about you, but i consider both 2.8 and 3.0 to be "close" to 3.1 -
>>as a physicist, i tend to think of numbers within 10% as equal ;-)
>
>For the default version it is 2.8 and that's far away from 3.1.
>
>10% is not 'a little bit'. In computerchess people like Frans Morsch work an
>entire year for 0.5%.
>
>Further Hyatt has done a very clear claim in journal of ICGA that it is
>
>0.7(n-1) + 1 for *any* number of processors.

Would you please cite the exact reference where I made that claim?

hint:  you won't find it in the JICCA +anywhere+.

In fact you won't find it as stated +anywhere+.

What you will find is this (and not in the JICCA):

quote on -------------------------------------------

speedup = 1 + (NCPUS - 1) * .7

which has been tested from on machines with from 1 to 8 processors.  For
machines with more than 8 cpus I have not done any testing to see how that holds
up.

quote off  -----------------------------------------

That has _only_ been published either here or in r.g.c.c...

Now is your big opportunity to show that your quote is correct.  Please first
cite the ICCA journal article where the above formula is given.  And then, for
your second challenge, cite _any_ post where I said the above was good for > 8
processors.  For 8 it works.  I have some 8-way opteron data to support it, but
then I have lots of 2-way and 4-way opteron data that my comments have been
based on in the past.

So either put up or shut up.  Cite the quote you claimed I made above.  Or else
retract it.





>
>This where at n=4 it is already dead wrong.

I sent you a complete log file for the CB / MchessPro test positions that were
run on my quad xeon.  The speedup was 3.0.

The formula predicts 3.1.

That is "dead wrong"???




>
>What about n = 64 ?



What about N=64?  I've never posted numbers beyond N=8.  I, unlike you, prefer
to post real numbers that come from real testing.



>
>So his entire article about crafty parallel scaling lineair with the number of
>cpu's getting 0.7 speedup out of any number of cpu's is just complete wrong
>again.

What "entire article" are you talking about?  I assume you are going to give a
citation since you keep mentioning it?  Note that a reference to "your head" is
_not_ a citation.  Journal name, issue, date, pages, title and author are
accepted as a citation.


>
>Crafty at his own machine doesn't even get 3.1. Instead it gets 2.8.

It got 3.0.  You have the log.  GCP got 2.8.  You have that log.  You _really_
have an ignorant streak as wide as the  Nile river.  There is no "one specific
speedup number.  It is a very non-deterministic number for all of us except you.
 It also depends on the particular position set used.  You treat it like it is
an absolute truth, when in reality you are only showing absolute ignorance...

But I _really_ don't believe you are that ignorant.  I believe you are that
dishonest to act like you are that ignorant however...


>
>Of course not coincidental this happens when someone else than Hyatt himself is
>doing the test at his machine.
>
>In general Hyatt takes positions where the mainline is directly trivial to find
>so easy move ordering possible and where therefore splitting is never going to
>suffer any problems. Note he tried to 'disproof' the 2.8 at 30 positions by
>showing 3 positions he analyzed himself, both having the same 'boring mainline'
>conditions.


I used the positions you asked for.  You said "Crafty on the DTS article
positions produces no speedup at all on a dual."  Remember that?  I ran the
numbers and got something like 1.6 but you noticed that my NPS was too low on
the dual, remember that?  Of course not.




>
>The word objective testing and statistical analysis are not in his dictionary i
>guess.

The words "honesty", "integrity", "scientific" are not in yours.  Perhaps we are
even???



>
>For the DTS article he has been proven frauding and he figured out a new formula
>to 'show' that his implementation lineairly scales (in speedup) at big machines.

Once more, please cite any article about DTS where there was a "linear" scaling.
 The DTS article definitely did _not_ show nor claim that.  It did great at 2-4,
not very good at 16.

That is not "linear" (which you can't seem to spell any better than you can use
it properly).



>
>Note that just for 4 processors the difference averaged over a lot of positions
>from 2.8 versus 3.1 is a *big* difference.

To you, perhaps...

>
>*all* tests done by others than hyatt show different numbers than what he writes
>up in journal of icga. then he complains that the machines tested are not good.
>
>then a test over 30 positions shows the point at his own quad and he still is
>disputing now because it falls 'within 10%' ?

Please cite the ICCA journal article you refer to before we go any farther.

Then we'll see just how far into the twilight zone you have ventured.  Note that
I _am_ workong on a crafty parallel search paper.  to date _nothing_ has been
published about it.  So I await with great interest your citation for the crafty
parallel search article in a JICCA issue dated in 2004 or earlier.  2005 and
beyond don't count as most of us can't see the future...

Your move...

>
>What a nonsense.

Do you ever have that right...



>
>
>
>>cheers
>>  martin



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.