Re: Probably been asked a hundred times before.

On Tue, 24 Jun 2008, Kevin Hunter wrote:

> Short of a response, I've read a number of reports that given some
> tuning FreeBSD 7.0 is the current top performer.

Those reports are all not quite right and I'm trying to get time to fully
debunk them in PostgreSQL land.

First off, they were running a small read-only benchmark, which is not
representative at all of real database performance. The FreeBSD team was
looking for something that stressed database kernel operations, and never
intended this to be a true database comparison.

Second, there was a problem with the new Linux CFS scheduler running
sysbench at the time the FreeBSD 7.0 reports touting its superiority were
released. It's since been fixed;
http://www.kernel.org/pub/linux/kernel/people/npiggin/sysbench/ shows you
the after and links to the before when using MySQL. PostgreSQL sysbench
results also benefitted, I haven't seen someone do a new set of benchmarks
there yet.

Third, there was also a bad interaction between the kernel and the
malloc/free sections of glibc that really impacted results here. The
FreeBSD 7.0 results had a specific fix in this area for their kernel.
Shortly afterward, a similar one was merged into Linux:
http://lkml.org/lkml/2007/4/1/268 (that fix was also active at the point
the previous benchmarks I pointed to were done)

So, yes, there was a brief window where the new FreeBSD 7.0 had a
performance advantage over Linux on this artificial (and bad) benchmark
running both its old scheduler and the still buggy and new CFS one, but
the two issues responsible have been resolved and current Linux kernels
using CFS are back to being on top again. The exact performance you'll
get depends on which Linux distribution/kernel combination you use, but
it's just not true that FreeBSD has an unambiguous lead here. The minute
the FreeBSD team declared that these benchmark results were somehow
interesting, it was simple for the Linux team to blow right by them by
optimizing for the weird things sysbench does the same way.

--
* Greg Smith gsmith(at)gregsmith(dot)com http://www.gregsmith.com Baltimore, MD