Hello, I need an advice about POP3 and NFS with Dovecot 1.0 (as 1.1 is still tagged beta) . I currently have all messages and indexes under a NFS partition . I thus turned cache off using actimeo=0 . Of course, my current architecture does not allow me to keep the same user on the same dovecot server. Now, I have some NFS issues with high percentages of getattr() . If I understood well, the cache must be disabled, because of the case of 2 differents computers could log on the same time and mess up the index file . If I use the directive pop3_lock_session and enable it, that should block conccurent pop3 access using a lock file . Can I consider it safe enough to turn on the NFS cache (ie removing actimeo=0) ? Thanks in advance Salim Ps: For you info, we use dovecot here for a big cable broadband ISP, with 2,000,000 mailboxes using 30 dovecot POP3 servers,10 dovecot proxy servers and we are really happy with your nice software, we expect a lot from 1.1 .
On Wed, 2007-11-21 at 22:08 +0100, salimlist at sdv.fr wrote:> If I use the directive pop3_lock_session and enable it, that should > block conccurent pop3 access using a lock file . > Can I consider it safe enough to turn on the NFS cache (ie removing > actimeo=0) ?What OS do you use? I guess with pop3_lock_session=yes you should be somewhat safe with Linux/Solaris, but not entirely. If you use BSDs then locking doesn't work correctly. I've been trying hard to figure out how NFS caches for for the last week and I think today's the first day I've finally understood it all correctly. :) I've been writing about those issues to http://iki.fi/tss/nfs-coding-howto.html With my newly found understanding even actimeo=0 isn't enough with v1.0, at least if your NFS server doesn't support nano/microsecond resolution for timestamps. And I should still do a few changes to v1.1's NFS cache flushing.. -------------- 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/20071122/dd3cbd81/attachment-0002.bin>
At 22/11/2007 04:19, Timo Sirainen wrote:>On Wed, 2007-11-21 at 22:08 +0100, salimlist at sdv.fr wrote: > > If I use the directive pop3_lock_session and enable it, that should > > block conccurent pop3 access using a lock file . > > Can I consider it safe enough to turn on the NFS cache (ie removing > > actimeo=0) ? > >What OS do you use? I guess with pop3_lock_session=yes you should be >somewhat safe with Linux/Solaris, but not entirely. If you use BSDs then >locking doesn't work correctly.I use Linux kernel 2.6.16.20 .>I've been trying hard to figure out how NFS caches for for the last week >and I think today's the first day I've finally understood it all >correctly. :) I've been writing about those issues to >http://iki.fi/tss/nfs-coding-howto.htmlThat a very useful document, I tryed on my side too to understand the linux cache, thanks !>With my newly found understanding even actimeo=0 isn't enough with v1.0, >at least if your NFS server doesn't support nano/microsecond resolution >for timestamps. And I should still do a few changes to v1.1's NFS cache >flushing..I think I will turn pop3_lock_session to on and enable the NFS cache to see the real benefits in term of NFS performance. If you are interested with results, let me know . I will also count the # of occurences of bad index files found in the log to see how safe that configuration is. Thanks a lot for your help . Salim