[Tfug] a comparison of file systems

[Brian Murphy]

Quoting Matthew Eskes <meskes at azcomputercentral.com>:
> Thanks Brian, I didn't know that. :)

Let me put in the standard disclaimer that XFS could have improved since
the last time that I used it. :)

I think the blogger linked below doesn't know that journaling can be
done at different levels.  XFS only journals filesystem metadata.  This
is explained in the last entry of the XFS FAQ
<http://oss.sgi.com/projects/xfs/faq.html>.

Other filesystems like ext3 will allow you to choose your level of
journaling.  ext3 uses the data= option for this.  Gentoo has a good
writeup at
<http://www.gentoo.org/doc/en/articles/l-afig-p8.xml#doc_chap4>.  I
don't know if they have answered the data=journal speed anomaly yet.

There is a Sun whitepaper out there that uses the postmark benchmark to
show that ufs kills ext3 and reiserfs in transactions.
<http://www.sun.com/software/whitepapers/solaris10/fs_performance.pdf>
It is flawed because if you look at the setup, the files written add up
to about 300MB on a machine that has 1GB of memory (think cache).

The paper gives a nice history and overview of ufs.  It mentions that
ufs only logs metadata.  There is also the warning of data loss after a
crash because the data isn't logged.  If I gather things from his blog
entry correctly, he was working with 30GB of data on a system with
512MB memory and possibly other activity besides the file cleanup he
was performing.  It's quite different from the Sun whitepaper scenerio.

The answer to his question why XFS isn't much slower than ext2 is that
XFS is only journaling metadata.  Why UFS is so slow when supposedly
only journaling metadata is something that I can't answer.

I don't know why some filesystems loose more files during an abrupt
outage either.  It could be related to disk cache on the physical disk,
filesystem on-disk layouts, flushing algorithms, fsck reconstruction
process, or ???  I'm not a filesystem internals expert.  Other hobbies
have distracted me from writing my own filesystem. ;-)

I'm surprised that he claims to have lost so much with reiserfs.  It
makes me wonder if he tried the --rebuild-tree fsck.reiser option.

I haven't read it all, but here is link to a presentation done on ext3:
<http://olstrans.sourceforge.net/release/OLS2000-ext3/OLS2000-ext3.html>
  It's excellent reading from the parts that I skimmed.

The same blog entry tells how the author is losing faith in linux
because of "exploits popping up like wildfire."  He goes on to mention
the 2.6.12, 2.6.12.1, and 2.6.12.2 kernel releases happening in about a
week's timeframe.  That's a bad example because out of those 3 releases,
only 2.6.12.1 was a security release.  Vendor kernels exist to rid you
of the upgrade-every-subpoint release cycle.

I happen to use ext3 almost everywhere because that is what Red Hat
installs by default (and it seems to work).  This relates to the
blogger's rant on linux needing time before it becomes "Unix".  The big
proprietary applications that I can think of only support certain
enterprise versions of linux.  They do this because enterprise distros
will take care of bugs (security or otherwise) by backporting the fix
to the version shipped with the distro.  The fineprint of proprietary
applications usually list a handful of patches necessary if you are
going to run an older solaris release.  He can't have it both ways...a
bleeding edge tweaker system and a proprietary friendly environment.
We have patched our propriety unix systems and needed to back out the
patch because it majorly broke functionality.  Nobody is perfect.

Solaris works but it definately has quirks of its own.  One of the more
bothersome ones is that a Solaris disk is not a Solaris disk.  You have
to look at the release date to be sure that it's compatible with the
hardware that you intend to install it on.  Perhaps linux gets a bad
rap because they increase the version number instead of calling it
version x released mm/yy.

Hope this helps!

Brian

> From what I've been reading as well, it seems that the Linux File systems
> don't seem to journal as well as UFS in Solaris. I haven't personally
> experienced this, so I intend to test this theory out when I can afford to
> build my next server box by installing Solaris 10. Again, I don't really
> know as of now, so I look forward to checking this out.
>
> Btw, this is one of the links that I have to refer to:
>
> http://www.neko-net.org/tiki-view_blog.php?blogId=3
>
> The linux which reads "Solaris is not Linux, really, and also not the other
> way around."
>
> Again, I haven't seen this for myself so any good conversation and pointer
> that y'all my have are very welcome, seeing how I'm heavily considering
> switching to Solaris.
>
> Matt
>
<snip>

The opinions or statements expressed herein are my own and should not be
taken as a position, opinion, or endorsement of the University of
Arizona.