Re: Benchmark (Phoronix): FreeBSD 9.0-RC2 vs. Oracle Linux 6.1 Server

Am 16.12.2011 08:06, schrieb O. Hartmann:
> For the underlying OS, as far as I know, the compiler hasn't as much
> impact as on userland software since autovectorization and other neat
> things are not used during system build.
> 
> From my experience using gcc 4.2 or 4.4/4.5 does not have an impact
> beyond 3% when SSE isn't explicetly enforced.

Well, but the compute intensive tests showed performance variance of a
few percents only, IIRC. The big differences were in the parts that
heavily depend on file system and buffer cache concepts (i.e. the low
limit on dirty buffers in FreeBSD, which is very beneficial in real
world situations; do you remember the first few releases of SunOS-4,
which heavily suffered in interactive performance due to a naive unified
buffer cache VM system that did not limit the amount of dirty buffers?
It caused interactive shells to be swapped out within seconds on systems
with background jobs writing to disk).

> More interesting is the performance gain due to the architecture. I
> think it would be very easy for M. Larabel to repeat this benchmark with
> a "bleeding edge"  Ubuntu or Suse as well. And since FreeBSD 9.0 can be
> compiled with CLANG, it should be possible to compare both also with
> "bleeding edge" compilers, say FreeBSD 9/CLANG, Ubuntu 12/gcc 4.6.2.

Clang may be considered "bleeding edge", but in quite a different way
than gcc-4.6.2. While the latter can look back on 2 decades of
development, clang is still in a state where feature completeness (and
bug-to-bug compatibility with GCC ;-) is much more important than
performance. there is much promise of powerful optimizations becoming
available in clang once it is mature, but just now expect GCC 4.6.2 to
deliver 5% to 10% higher performance than clang.

But as stated before: To exclude compiler dependencies just run the
Linux binaries on FreeBSD. There is slight emulation overhead and Glibc
is not particularly optimized for FreeBSD, but this will still provide
more useful results.

And the tests should be selected to represent reasonable real-world
scenarios. Server programs tested on otherwise idle systems and running
for just a few seconds (not reaching equilibrium during the majority of
the test period) are not representative at all (again: if your goal is
to compare server performance).

Regards, STefan