Re: Performance hit with --enable-atomic-refcount

[John W. Eaton]

On 06/06/2014 01:40 PM, Rik wrote:
> 6/6/14
>
> All,
>
>  From the previous benchmarking, there was a doubling in execution time from
> 3.8.1 to the gui-release branch for a nested for loop.  It turns out that
> this doubling is 100% correlated with using the --enable-atomic-refcount
> option to configure.  I built the gui-release branch with this feature
> disabled and the results are then equivalent to the 3.8.1 release.  It
> seems like we need to explore a different solution than this configure
> option so that we can both use Qt graphics and have reasonable performance.

Here are the results I see.  I used default configure options in both 
cases except for specifying --prefix and --disable-atomic-refcount in 
one build.  I also used cputime instead of tic/toc, but that shouldn't 
be too important if your system is mostly idle when you run the tests.

  with atomic refcount:     ~ 1.60 CPU seconds
  without atomic refcount:  ~ 1.25 CPU seconds

So I see a performance hit, but it is not a factor of 2.  I remember 
that we had some significant differences between our results the last 
time we looked at this, but the test then was running the test suite, 
not just a single nested for loop with an addition expression.

I'm using gcc version 4.8.2 (Debian 4.8.2-1) on an amd64 system.  I 
didn't try to fix the CPU frequency as I'm not sure how to do that.

jwe