Timo Sirainen skrev:
On Wed, 2009-10-14 at 21:14 +0200, Mikkel wrote:
It has been my wish to move to dbox for a long time hoping to reduce the number of writes which is really killing us. BTW. Have you tried maildir_very_dirty_syncs=yes setting? That should reduce disk i/o, and I'm really interested in hearing how much.
I don't think I've tried that one. Earlier on I experimented with fsync_disable=yes (which made a huge difference by the way) but that was before I started using mail_nfs_storage=yes and mail_nfs_index=yes
I would like to try using maildir_very_dirty_syncs=yes but is it advisable in combination with NFS?
It should be fine with NFS if indexes are also on NFS. Although I just fixed a bug related to it: http://hg.dovecot.org/dovecot-1.2/rev/7956cc1086e1
The system is currently running dovecot version 1.1.19. Would you consider it safe to try it on that version as well?
Indexes on NFS are problematic now though if multiple servers can access the mailbox at the same time. mail_nfs_index=yes is supposed to help with that, but it's not perfect either. Long term solution would be for Dovecots in different machines to talk to each others directly instead of through NFS.
Is worse now than previously?
I have been running at production setup with two servers accessing the same Maildir data from NFS without any problems for quite a while now. Load is spread randomly between the two servers so I can only assume that by coincidence they sometimes try to access the same mailbox. This has functioned quote well with many versions of the 1.1.x dovecot releases so unless some new issues have been introduced I don't think I should fear anything in that regard :-)
Regards, Mikkel