Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Crap statement refuted about parallel speedup (part 2)

Author: Robert Hyatt

Date: 21:29:22 09/21/01

Go up one level in this thread


On September 21, 2001 at 20:53:37, Vincent Diepeveen wrote:

>>
>Note that some positions also Crafty goes
>faster than 2x  at 2 cpus. Did you forget that?
>
>Now imagine that a few more positions this would be the case, so
>you're missing the boat somewhere!!!!!!!
>
>My advice, hop in and give it a try!
>


I always try.  But I also know that I am not going to search more than 2x
faster with two cpus no matter how hard I try, except for the anomalies
that happen on rare occasions.




>There is another thing which needs to be mentionned when talking
>about 4 cpu's or more: that's your load balancing is having a worstcase
>speedup from 1.0, which is not a very promising idea. At 4 cpu's that
>means that you might use it 25% effective in a worst case scenario.
>
>At 2 cpu's that is 50%.
>
>So 2 cpus is an optimal number of cpu's to run parallel at IMHO.


That is nonsense.  And it is a totally unsound way to think about this.
You are saying you want to maximize the worst case, by drastically hurting
the "normal case".  That makes absolutely no sense.


>
> a) the worst case is better than that of any other number of processors
>    other than n=1
> b) you can profit from parallel luck in cutoffs, move ordering improvements
>    etcetera.
>
>Why did i never hear *anyone* about the worst case
>speedup its parallel algorithm is going to give?

Maybe because you don't read everything?  _I_ certainly covered the worst-cases
in the JICCA journal.  And I have talked about one position that used to wreck
me in a set of positions Bruce and I ran to compare our parallel searches.




>
>I mean there are pretty easy load balancing ideas. One of them is
>one i'm using myself. That's simply demanding that each splitted leaf
>gets 50% of the processors.
>
>So expected worst case is 50% (though there are theoretical possibilities
>that it's worse than 50%, the likelyness of that is way way smaller than
>with the way crafty is running parallel).
>
>This where expected worst case in crafty is 25%.

If I thought your idea was good, I would be doing it.  Crafty _already_ has
code to limit the number of processors at a split point.  I currently have it
set to 4, which seemed to be best for me when I was testing on 16 cpus.  I
don't think 2 is better although anybody can test this.

smpgroup=2 in the crafty.rc file will change this.  I have tried it and it
didn't work as well as 4 in the test suites I ran.  Bet you didn't know this
was _already_ there did you?  :)

Here's a bit of something to think about.  I'll bet that if you have
thought about it, I have already tried it.  I started this SMP stuff a
_long_ time before you did.  16 years or so earlier in fact, since my
parallel search was playing chess in 1983.

Don't assume _everybody_ is an idiot.  Don't assume that because you have
studied something for a year, that you know everything about it.  Both
assumptions are really bad...

If you want to send your latest here and have me run some tests to compare
our performance on my quad, I'll be happy to run the 1 vs 4 tests for a group
of test positions (say the 24 kopec positions, or the positions Bruce uses).

That way we can compare our searches to see who is getting what, when...





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.