NFS is commonly used in one of these ways:
- Dovecot is run in a single computer.
- Dovecot is run in multiple computers, users are redirected more or less randomly to different computers.
- Dovecot is run in multiple computers, each user is assigned a specific computer which is used whenever possible.
The only way to reliably implement the 2nd setup is with the director service.
Single Dovecot server setup or Dovecot director cluster setup:
mmap_disable = yes #dotlock_use_excl = no # only needed with NFSv2, NFSv3+ supports O_EXCL and it's faster mail_fsync = always mail_nfs_storage = no mail_nfs_index = no
Multi-server setup that tries to flush NFS caches (increases NFS operations, and isn't fully reliable), try not to use this:
mmap_disable = yes #dotlock_use_excl = no # only needed with NFSv2, NFSv3+ supports O_EXCL and it's faster mail_fsync = always # These settings slow things down and don't fully work, use director proxy instead: mail_nfs_storage = yes mail_nfs_index = yes
Run ntpd in the NFS server and all the NFS clients to make sure their clocks are synchronized. If the clocks are more than one second apart from each others and multiple computers access the same mailbox simultaneously, you may get errors.
NFS caching problems
NFS caching is a big problem when multiple computers are accessing the same mailbox simultaneously. The best fix for this is to prevent it from happening. Configure your setup so that a user always gets redirected to the same server (unless it's down). This also means that mail deliveries must be done by the same server, or alternatively it shouldn't update index files.
Dovecot flushes NFS caches when needed if you set mail_nfs_storage=yes, but unfortunately this doesn't work 100%, so you can get random errors.
Disabling NFS attribute cache helps a lot in getting rid of caching related errors, but this makes the performance MUCH worse and increases the load on NFS server. This can usually be done by giving actimeo=0 or noac mount option.
If you keep the index files stored on NFS, you'll need to set mmap_disable=yes. If you're not running lockd you'll have to set lock_method=dotlock, but this degrades performance. Note that some NFS installations have problems with lockd. If you're beginning to get all kinds of locking related errors, try if the problems go away with dotlocking.
With mbox/Maildir format it's also possible to store index files on local disk instead of on NFS. If the user gets redirected to different servers, the local indexes are automatically created/updated. If the user is (nearly) always redirected to the same server this should be fine and you would likely get higher performance than indexes stored on NFS, but if the server changes it can be slow to recreate the index/cache files.
Single computer setup
This doesn't really differ from keeping mails stored locally. For better performance you should keep index files stored in a local disk.
Random redirects to multiple servers
You should avoid this setup whenever possible. Besides the NFS cache problems described above, mailbox contents can't be cached as well in the memory either. This is more problematic with mbox than with maildir, but in both cases if a client is redirected to a different server when reconnecting, the new server will have to read some data via the NFS into memory, while the original server might have had the data already cached.
If you choose to use this setup, at the very least try to make connections from a single IP redirected into the same server. This avoids the biggest problems with clients that use multiple connections.
Per-user redirects to multiple servers
This method performs a lot better than random redirects. It maximizes the caching possibilities and prevents the problems caused by simultaneous mailbox access.
New mail deliveries are often still handled by different computers. This isn't a problem with maildir as long as you're not using LDA (i.e. dovecot-uidlist file or index files shouldn't get updated). It shouldn't be a problem with mboxes either as long as you're using fcntl locking. This problem can be fully solved by using LMTP protocol to deliver the mails to the correct server (possibly using Dovecot's LMTP proxy).
Here's a list of kernels that have been tried as NFS clients:
FreeBSD has a caching bug which causes problems when mailbox is being accessed from different computers at the same time
Linux 2.6.16: utime() is buggy, fix in here. With the fix applied, utime() seems to work perfectly. High-volume systems may experience VFS lock sync issues and for these the complete patchset at http://www.linux-nfs.org/Linux-2.6.x/2.6.16/linux-2.6.16-NFS_ALL.dif is suggested and appears to work well in production.
- Linux 2.6.18: Seems to have intermittent caching issues. The same .config with 22.214.171.124 has been tested and appears to work well.
- Linux 2.4.8: Has caching problems, don't know if they can be solved
Solaris: If it's completely broken, see http://dovecot.org/list/dovecot/2006-December/018145.html
The Connectathon test suite is very useful to verify a healthy NFS setup, see http://www.connectathon.org/nfstests.html
- readdirplus isn't really needed by Dovecot and it can slow down some NFS servers. Use "nordirplus" mount option to disable it.
- Dovecot doesn't care about root_squash setting, all the root-owned files are in /var/run typically which is not in NFS
In an environment using Debian (2.6.18) clients with Isilon NFS cluster nodes - the following mount options were found to be the most successful: rsize=32768,wsize=32768,hard,fg,lock,nfsvers=3,tcp,retrans=0,nordirplus 0 0
To learn more about NFS caching and other issues, mostly from a programmer's point of view, see NFS Coding HOWTO
- Use such permissions for the unmounted mount point root directory that Dovecot can't create files under it. Otherwise if the NFS server isn't mounted for any reason and user access mails, a new empty user mail directory is created, which breaks things.