[Dovecot] v2.0.6 released

Stan Hoeppner stan at hardwarefreak.com
Tue Oct 26 01:42:56 EEST 2010


Timo Sirainen put forth on 10/25/2010 10:25 AM:
> On Fri, 2010-10-22 at 06:02 -0500, Stan Hoeppner wrote:
>> Timo Sirainen put forth on 10/21/2010 4:10 PM:
>>> On 21.10.2010, at 19.14, Timo Sirainen wrote:
>>>
>>>> 	+ mdbox: Added mdbox_preallocate_space setting (Linux+ext3/XFS only)
>>>
>>> ext4 of course, not ext3. Also requires Linux kernel v2.6.23 or newer and glibc 2.10 or newer.
>>
>> Is this available for mbox or just mdbox?  If mdbox only, why not mbox also?
> 
> It's only for mdbox because they have a known upper size
> (mdbox_rotate_size). With mbox the only possibility would be to let
> admin specify how large the files should be preallocated to. And how
> would the admin know? And should different mailboxes have different
> preallocated sizes? For example "Drafts" mailbox is unlikely to ever
> grow very large. And in any case, if you care about performance mbox
> isn't really the best choice.

Is it still the best choice performer WRT full text search?  I use fts
squat, but it seems whenever I do a body search of an IMAP folder that
the index is cold, thus I don't often get the benefit of the index.
Would maildir, dbox, or mdbox give better performance here with a cold
index?  Full text search is the main reason I've stayed with mbox.  I've
never noticed any performance issues (not after you recommended I enable
very_dirty_syncs that is).

That's all kinda off topic though.  The main reason I asked about
preallocation in first place was merely to limit file fragmentation.  As
I understand it, some types of preallocation can be each time the file
is written, in essence extending the tail at the end of the file.
Something like this won't eliminate fragmentation, but it should lessen
the number of fragments as the file grows.  I'm not currently running
into performance problems due to fragmentation, but it could/will happen
at some point.  Currently I'm running xfs_fsr twice a week via cron to
keep fragmentation at bay.  Those running mbox on any other filesystem
would benefit from this limited fragmentation as no other FS has a
working online defragmenter.

Anyhow, it sounds like even if it could somehow be done with mbox it
would be pretty difficult to implement, and not provide much ROI as
apparently few OPs still use mbox.

It's really cool to hear that preallocation should work very well WRT
eliminating fragmentation of mdbox as you said their is a maximum file
size.  This seems like a perfect fit.

-- 
Stan


More information about the dovecot mailing list