Tuning Postgresql on FreeBSD 5.1
tlambert2 at mindspring.com
Wed Aug 27 03:27:26 PDT 2003
Maciej Freudenheim wrote:
> > issues (if for no other reason than UFS+softupdates). Linux has
> > problems with memory over commit situations and file system integrity
> > while still maintaining reasonable file system performance (ext2 is
> > faster than ext3 by a wide margin, but ext2 is _not_ a reliable FS).
> Even heard of reiserfs or xfs?
> I'm not going to start Linux advocacy here, but it's so funny for me
> when somebody says 'leenoox sux, it has crappy filesystem' that i can't
> leave it without reply :)
He didn't say that. He said that it had problems with memory
overcommit situations; most OS's do. FreeBSD has done a lot
of work on graceful degradation (i.e. it doesn't crash when it
runs out of memory and there's no more swap available). So it's
you whose bringing that interpretation to the data.
I haven't seen any reasonable benchmarks on ReiserFS or XFS vs.
Ext2FS or Ext3FS that didn't involve sticking too many files in
a single directory and actually measuring btree vs. linear
directory entry layout instead of actual raw I/O performance.
For most modern hardware, you never get CPU bound, so the real
determining factor on raw I/O performance is almost always going
to be raw disk I/O speed, and safe speed is going to be lower
than the manufacturer benchmarks; some disks are "unsafe at any
speed", to paraphrase Ralph Nader, because they cache when you
tell them not to using the commands the manufacturer has stated
actually work to tell them not to.
Also, FYI, Postgress, unlike Postfix or qmail, doesn't dump a
zillion files into the same directory so they can measure
themselves against a situational benchmark; they are actually
smart enough to port to the POSIX interface, and not depend on
FS implementation-specific tricks to get their performance.
If you wanted to join the XFS porting project, I'm sure they'd
like the help. If you wanted to port ReiserFS, realize that
software patents will probably prevent your work from being
used in the U.S. or Japan, unless you could prove it to be
non-infringing on the Novell D.O.W. patents.
More information about the freebsd-performance