Cyrus 2.3.8 thanks for help

[Bron Gondwana]

On Wed, 12 Sep 2007 16:57:48 -0700, "Vincent Fox" <vbfox at ucdavis.edu> said:
> We had several other universities called and offered assistance.
> 
> Thanks!
> 
> We also had several people call with offers to help with consulting
> on it, but mostly not at the same level of Cyrus hardware/software/config
> we are running.  We appreciate the effort though, we really do.  Lots of
> folks seem to run mail-stores in the 5K-10K user range, not too many
> trying to stuff 30K users onto one that we have talked to so far.

We used to - now we stuff about 10k users per cyrus instance, but multiple
instances on the one machine.  Makes load balancing and failover much more
sane - and we can recover an instance in about a day, which means that
we're not running single-replica for longer than necessary.

> We are still struggling with some slowness on the system and trying to
> refine our setup.   Probably we are trying to do too much at once.

Possibly.  Do you have split meta?  That could possibly help, though I
don't know how much difference it makes on ZFS.

> Our setup:
> Cyrus 2.3.8 + FastMail patches
> Sun T2000, 16 gigs RAM, Solaris 10u3
> 2 3510FC arrays using Sun MultiPath
> ZFS mirrored pools across the 2 arrays (expensive but safe)
> ZFS snapshots for 14 days

Pretty funky.  We're only using ZFS on one machine, and that's our
backup server which has totally different usage patterns than an IMAP
server.

> We have cut down LMTP maxchilds to 200 then 150 seems to
> be helping with locking issues.  I think we also have other issues
> such as doing sendmail on the mail-stores instead of LMTP, and
> we realize now we compiled SASL with /dev/random.  Ken replied
> that we should turn up logging to DEBUG haven't found where to
> do to that just yet hopefully tomorrow.

syslog config.  We have:

local6.info                                             -/var/log/imapd.log

You probably want:

local6.debug                                            -/var/log/imapd.log

Or something like that.

> Hope this follow-up helps other list-readers.

I'm afraid we don't have the experience with ZFS to help you tune there.
We run reiserfs on Linux systems.  I suspect that ZFS could well be the
performance culprit.  We haven't found any filesystem quite as good as
reiser for the usage pattern we produce (though we're currently in an
email exchange with various reiserfs experts about some strange memory
use/caching behaviour - a quote "the FastMail guys find all the interesting
bugs" :)

Bron.

-- 
  Bron Gondwana
  brong at fastmail.fm