Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Here are some actual numbers

Author: Robert Hyatt

Date: 07:54:50 04/14/03

Go up one level in this thread


On April 14, 2003 at 05:28:11, Sune Fischer wrote:

>On April 13, 2003 at 11:15:15, Robert Hyatt wrote:
>
>
>>Absolutely no idea.  Eugene posted that he has an application that speeds up
>>by exactly 2.0 when turning SMT on.  I have gotten 20-30% improvements in NPS,
>>and 0-20% improvements in raw time-to-solution.  I posted the files needed to
>>run the test yourself.  I ran them a total of six times, and things were
>>pretty consistent.
>
>These numbers fit as far as I can tell, but I do believe there can be cases
>where the parallel search overhead is so big, that even a 30% boost from SMT
>won't be enough to overcome it(?).

This can happen.  But hopefully it won't happen on a consistent basis.
Otherwise the parallel search algorithm used is simply no good.  IE if you
write something that is consistently _slower_ with two processors, then it
is not very useful.


>
>Have you tried investigating when SMT is no longer worth it?

I don't see any cases where it is not worth it.  I'm sure there is the
occasional position where a two-thread parallel search is no faster than a
single thread, and the very rare case where it is actually slower.  But the
great majority of positions run faster with two threads than with one, assuming
there are two cpus to use.  At least for Crafty.

If you look at the SMT data I posted, using four positions Vincent wanted to
see, three posted clean SMP speedups when going from two to four threads.  One
posted a slower time using four instead of two, for one run out of three, with
the other two being just barely above a break-even point (not much speedup.)




>Say you had a 4 way box with SMT (8 threads). Wouldn't it be faster in
>time-to-solution to just use the 4 threads, even if using 8 gave you a 30% nps
>increase?


Not unless the parallel search overhead is either very high, or is non-linear
with the number of cpus.  I've run on SMT quads and the extra "four" logical
processors definitely help.  Being from 10-20% faster than using four threads
with SMT disabled.  Which matches my dual numbers pretty well.

I've not run on an 8-way or beyond that had SMT however so I can't say that it
will work for all boxes.




>
>-S.



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.