DINH Viêt Hoà a écrit : > On 6/27/07, damien chambe - EGS <[EMAIL PROTECTED]> wrote: >> Timo Sirainen a écrit : >> > On Fri, 2007-06-22 at 00:28 +0200, damien chambe - EGS wrote: >> > >> > >> >> Our Dovecot server had to be changed due to a hardware problem >> >> I was forced to use SUSE SLES 10 for the new one, instead of SLES >> 9 on >> >> the old server. >> >> >> > >> > The kernel matters a lot with NFS. Some kernels are more broken than >> > others. Attribute cache also matters. http://wiki.dovecot.org/NFS >> > >> > >> >> I store mails on NFS, and index on local disk. There's only one >> >> dovecot >> >> server, so no multiple access. >> >> >> > >> > So deliver is also run on the same server? If all of it is done on the >> > same server, then pretty much the only thing you can change is the >> > kernel or somehow try to work around its bugs. I can think of only >> this >> > fix on Dovecot's side: >> > >> > http://dovecot.org/list/dovecot/2006-December/018145.html >> > >> > Hmm. Actually I just realized another reason that could cause >> these: Are >> > the clocks on the NFS server and on your Dovecot machine synchronized? >> > They must be less than 1 second apart at all times or you'll begin to >> > see problems. >> > >> > >> I've tried to update kernel (SUSE SLES 10 is 2.6.16) but no change. >> But synchronizing NFS server and dovecot server with NTP did the trick. >> >> SUSE SLES9 NFS was more tolerant than SLES 10 with time sync... >> >> No more uid messages yesterday ! > > did you use the following on the NFS server : > option "no_subtree_check" in /etc/exports > I know that it solved some bugs when accessing files through mmap() on > a NFS filesystem. >
I can't easily modify export option on the NFS server, it's a Lifekeeper DRDB cluster, I've used noac for the mount on the dovecot side. It slows downs a little but it is very reliable on my config. Damien
