On 27.10.2010, at 15.48, Andre Nathan wrote:
On Tue, 2010-10-26 at 23:52 -0200, Andre Nathan wrote:
On Tue, 2010-10-26 at 22:39 +0200, Timo Sirainen wrote:
Have you set mmap_disable=yes?
Yep.
Just out of curiosity, is this setting really needed, or is it for performance reasons? OCFS2 claims to support mmap:
If mmap is supported, it's not necessary, but might still be better for performance (or might not, depending on filesystem's mmap implementation..)
This really seems to be related to the shared accounts. Our load balancer makes no attempt to send the same users to the same server; it just sends connections based on server load. This allows two simultaneous connections to access an account, one on each server. Given a distributed FS, this should be OK, right? At least in theory it's no different than two CPU cores accessing the same account in a single server.
In theory sure.. But Dovecot's indexes have already been heavily stress tested in local filesystems with multiple CPU cores and it should be stable nowadays. If you can easily break it with a clusterfs, the problem is clearly the clusterfs itself.
You could easily try this with imaptest http://imapwiki.org/ImapTest by running it for the same mailbox in two OCFS2 nodes at the same time and see if Dovecot immediately starts logging errors.