Re: pre-2.18 performance data

["Carlos O'Donell" <carlos at redhat dot com>]

On 07/20/2013 04:57 PM, Arjan van de Ven wrote:
> Hi,
> 
> [not sure if this sort of thing is appreciated on this list; if not let me know]

Arjan,

Good or bad this kind of data is very much appreciated and
very valuable, particularly coming from an objective 3rd
party.

> Yesterday I put an early snapshot of glibc 2.18 into my distro (using
> git hash 19a601f7b2fc69201c296f0720459a32dda5f4c3), to give it a spin
> for compatibility and performance.
> No compatibility issues so far;
> on the performance side, the good news is that there are no
> performance regressions in the tests that I run automatically, and
> there are some real performance improvements.
> 
> The "smallpt" benchmark (floating point heavy) is showing high-double
> digit %age improvements
> 
> XZ decompression, n-queens and the libxml2 performance tests all show
> low single digit %age improvements
> 
> a few tests are noisy and will get run a few more times in the next 24
> hours before any conclusion can be made.
> 
> graphical results are at http://linux.fenrus.org/performance/ (the
> datapoints prior to July 19th are with glibc 2.17, the data points for
> the 19th and later are with the above mentioned commit as system
> glibc)

During 2.18 the community worked hard to put in place a microbenchmark
testsuite that would allow us to objectively evaluate performance for
glibc patches.

Particularly for libm we've made sure that the patches showed
microbenchmark gains, and I'm happy that it shows.

The counterpoint to all this good news is that in 2.15/16/17 we 
consciously chose to fix conformance issues at the cost of performance 
issues.

It has only been with 2.18 that we've slowly optimized back up to the
level of our previous performance.

So while I'm happy to see that we're doing well I think you caught
us at a particularly good time to measure relative gains :-)

Cheers,
Carlos.