Very poor ext3 performance with dir_index, and reiserFS question

[Paul Dekkers]

Hi,

I did some filesystem-tests with some typical (real) mail-data to help 
us reconsider our filesystem-choice for our mail-servers.
I was a bit surprised by the bad performance of ext3 with the dir_index 
option enabled.

It seemed that writing to partitions with this flag enabled where 
considerably faster (about 40%, it was as fast as reiserFS now ;-)) but 
that reads where terribly slower: about 10 times worse then without this 
option!

Is this known behavior?
If it is; is my conclusion that ext3 without this option would be better 
true?

I put a small report of my findings and some graphs on:
http://www.surfnetters.nl/paul/fs/

I'm afraid I already know the answer, but my careful conclusion was that 
reiserFS is a better choice for this kind of application. Is there any 
(official) way to use reiserFS with RH enterprise (4)?

I tried to recompile the kernel package myself and enable reiserFS in 
that, and of course that worked (but this is a bit "houtje-touwtje" as 
we call that in Dutch, you probably get the point ;-)). And I still have 
to get the reiserFS utils, which I might just take from fedora's SRPMS, 
but...
(I tend to say that I'd rather do this then use ext3 with it's strange 
behavior (see P.P.S.))

Any advise or feedback is appreciated  :-) 

Regards,
Paul

P.S. I did this filesystem tests with Fedora 3, the platform I would run 
the mail-platform on would be Redhat 4 - I suppose/hope this didn't make 
a real big difference (except for the fact that reiserFS was available!) 
since RH 4 was based on FC3... more details on the hardware / software 
are in the report.

P.P.S. I also discovered that disabling dir_index with tune2fs and 
re-scanning with fsck didn't get the performance back again. So there 
you have a filesystem that appears to be without dir_index, but still 
has the terrible performance of ext3 -with- the option enabled. Right. 
I'd really vote for reiserFS.