[Dovecot] linux 2.4 vs 2.6 kernel

[Matthias Andree]

vmstech <vmstech at tpg.com.au> writes:

> ReiserFS does seems just like the ducksnuts in terms of performance, but
> it seems to have accumulated some horror stories along the way:
> http://www.linuxsa.org.au/pipermail/linuxsa/2002-January/038035.html
>
> Marcs being runing it for 5 years, no issues - any one else using it?

I'm currently phasing reiserfs out everywhere. While I don't recall
major accidents (should I write incidents? nevermind) since the later
3.6 versions in the later 2.4 kernels, reiserfs has its
limitations. Tail merging, space efficiency, is slow to write, so you'll
usually mount -o notail. reiserfs has a limited number of files per hash
value that a file system can hold, and the earlier reiserfs versions
became extremely slow in the face of hash collisions.  The quality of
the reiserfsprogs is IMHO inferior to e2fsprogs.

ext3 has learnt the dir_index feature (dirhash) a while ago, it's a
rather smooth migration (unmount, tune2fs, e2fsck -f -D to create
indexes, mount) - but it turns out I don't have directories that are
*so* full it would really matter for me. It appears to work well though.

> It also seems a non-trivial to convert an existing ext3 install.

There is no conversion. Backup, reformat, restore.

> If I'm going to go to that effort, I should probably contemplate xfs and
> jfs and make best choise first up - anyone got any experience with these
> where maildir is concerned?

I haven't dared put my Maildirs into either of these.

I used XFS once for a CVS spool and after a crash ended up with zero
blocks in the ,v files, which caused the cvs server to crash. I'm not
sure if that has been fixed since. I've never seen this happen with ext3.

SUSE are somewhat discontinuing JFS support in their SUSE Linux 9.3
distribution for "technical reasons" that aren't detailed.

Note however that ext3fs supports data journalling, so you can have all
the expensive scattered synchronous writes (and the Maildir format has
lots of these) turned into linear synchronous writes, which speeds up
things quite a bit.

-- 
Matthias Andree