Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Crap statement refuted about parallel speedup

Author: Robert Hyatt

Date: 20:11:47 09/21/01

Go up one level in this thread


On September 21, 2001 at 20:34:35, Vincent Diepeveen wrote:

>On September 18, 2001 at 23:30:51, Dave Gomboc wrote:
>
>I already knew there are better move orderings possible because
>i only get 0.85-0.90 cutoff rate versus most getting 0.90, however
>static move ordering code is hell harder to write than simply
>splitting in parallel and hoping it gives a quicker cutoff.
>
>The 0.92 from crafty is not entirely true , because we should not
>forget that its fliprate is 0.05 versus mine is under 0.01

I don't follow that at all.  If my 92% is higher than yours, then my
"fliprate" must be lower...  I don't see how it can be any other way.
But in any case, my statistics on splitting at the wrong nodes suggests
that I do very well there...





>
>So the qsearch can improve bigtime from crafty, because basically
>that causes this bad fliprate!
>
>That all our move orderings can be improved was already known
>and some years ago i posted bigtime about this subject but the same
>Bob wrote here:
>
>  "our move ordering is that optimal that we cannot improve it anymore".
>

I didn't say that.  I said that our move ordering is producing the right
move over 90% of the time, as I have carefully measured in my search
statistics.  92% is a good average with 95% pretty common.  That remaining
5% is not a huge problem.

I certainly stand by the statement that if you can produce a > 2 speedup using
2 cpus, then you can copy that algorithm to a one processor machine and
run faster than you have been running.  Because your one-processor algorithm
is bad.





>completely contradicting with what is getting said now!
>
>>On September 18, 2001 at 18:27:01, Robert Hyatt wrote:
>
>>>Before you label a statement as ridiculous, find me _one_ person that will
>>>side with you and say that "yes, it is possible to get a > 2 speedup using
>>>only two cpus on anything but anomaly positions.  Find _one_ person that will
>>>agree...
>>
>>This is (okay, not very!) amusing... the statement is the exact opposite of
>>reality, which (as Bob already knows) is that "it is only possible to get a < 2
>>speedup using only two cpus on anything but anomaly positions".
>>
>>If a parallel algorithm consistently outperformed a sequential algorithm, then
>>you've just discovered a better sequential algorithm as well (use time-slicing).
>> If you then don't use this better sequential algorithm to compare your parallel
>>algorithm against, you'd be comparing a good parallel algorithm against a shitty
>>sequential algorithm, which would make the speedup result worthless.
>>
>>It is absolutely key that when people compare their parallel algorithm to their
>>sequential algorithm that they compare the best possible sequential algorithm.
>>There are more than a few papers that don't do this... thankfully, at least some
>>of them have been rejected.
>>
>>Dave



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.