We are seeing some poor performance recently that is focused around users with large mailboxes (100,000 message /INBOX, 80,000 message subfolders, etc). The performance problem manifests as very high system% utilization - basically iowait for NFS. There are two imap servers with plenty of horsepower/memory/etc. They are connected to a 3050c cluster via gig-e. Here are the mount options: rw,tcp,nfsvers=3,rsize=8192,wsize=8192,nosuid,nodev,soft,intr,noac,actimeo=0 I need to get this system% utilization brought down because it is starting to slowdown imap performance for other users (it goes up to 95% on a regular basis, with users that have folders with many messages). My thought is to try removing the noac option and bump up the actimeo value from 0 to 30. Unfortunately I don't have access to use a SLB to direct traffic to individual servers. My understanding is that the worst that can happen is index corruption (which will result in rebuilding them when it happens). Any thoughts? I'm running RC24.
bofh list wrote:> We are seeing some poor performance recently that is focused around > users with large mailboxes (100,000 message /INBOX, 80,000 message > subfolders, etc). > > The performance problem manifests as very high system% utilization - > basically iowait for NFS. > > There are two imap servers with plenty of horsepower/memory/etc. They > are connected to a 3050c cluster via gig-e. Here are the mount > options: > rw,tcp,nfsvers=3,rsize=8192,wsize=8192,nosuid,nodev,soft,intr,noac,actimeo=0 > > > I need to get this system% utilization brought down because it is > starting to slowdown imap performance for other users (it goes up to > 95% on a regular basis, with users that have folders with many > messages). > > My thought is to try removing the noac option and bump up the actimeo > value from 0 to 30. Unfortunately I don't have access to use a SLB to > direct traffic to individual servers. > > My understanding is that the worst that can happen is index corruption > (which will result in rebuilding them when it happens). > > Any thoughts? I'm running RC24. >What operating system are you using? Performance over NFS is tremendously different. And each OS can be tweaked differently. You should try noatime (noxattr on solaris) regardless. Cheers
On 3/22/07, Tom Bombadil <grlists at gmail.com> wrote: ...> What operating system are you using? Performance over NFS is > tremendously different. And each OS can be tweaked differently. > > You should try noatime (noxattr on solaris) regardless.RHEL4u3 64-bit. I will try noatime. Thanks
We used to have RHEL4 talking NFS to bluearc (instead of netapp). Switching to solaris 10 improved performance about 10 fold. The app was postgres though, not dovecot. Cheers bofh list wrote:> On 3/22/07, Tom Bombadil <grlists at gmail.com> wrote: > > ... > >> What operating system are you using? Performance over NFS is >> tremendously different. And each OS can be tweaked differently. >> >> You should try noatime (noxattr on solaris) regardless. > > RHEL4u3 64-bit. I will try noatime. > > Thanks >
On Thu, 2007-03-22 at 14:19 -0400, bofh list wrote:> We are seeing some poor performance recently that is focused around > users with large mailboxes (100,000 message /INBOX, 80,000 message > subfolders, etc). > > The performance problem manifests as very high system% utilization - > basically iowait for NFS.Did you use Dovecot's quota plugin? With what backend? -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 189 bytes Desc: This is a digitally signed message part URL: <http://dovecot.org/pipermail/dovecot/attachments/20070326/31add39d/attachment.bin>
On 3/25/07, Timo Sirainen <tss at iki.fi> wrote:> On Thu, 2007-03-22 at 14:19 -0400, bofh list wrote: > > We are seeing some poor performance recently that is focused around > > users with large mailboxes (100,000 message /INBOX, 80,000 message > > subfolders, etc). > > > > The performance problem manifests as very high system% utilization - > > basically iowait for NFS. > > Did you use Dovecot's quota plugin? With what backend?We are not using quotas. As much as I would like to, it's not an option at this time.