So the load balancer is randomly assigning the user to different servers? Not really the preferred setup. I guess you've read http://wiki.dovecot.org/NFS? Yes, in our previous setup we just used round robin and it seemed to be fine, never had load problems (using courier). After switching to dovecot I had tried both, while trying to narrow down the problem to the pop3 I had set only 1 machine to handle all pop3 and I got a call the next morning from support and the load was 511.
login_processes_count: 15 login_max_processes_count: 256
This problem and possibly others could maybe be solved by using the high performance mode as described by http://wiki.dovecot.org/LoginProcess
I tried this the same night that I had a single pop3 box, it ran out of processes I think was what the log said, so I switched back to the current settings.
mail_location: maildir:~/:INDEX=/dovecot-index
Are index files on local disk or on NFS?
Yes, although I set INDEX=MEMORY for pop3 section after reading that on one of the google search results
mail_save_crlf: yes
This adds some extra disk I/O, probably better to set to "no" unless you have a good reason.
I forgot why I had enabled this, I just turned it off though.
mmap_disable: yes mail_nfs_storage: yes
If indexes are on NFS, you should have mail_nfs_index=yes too.
I had this enabled as well until I read the index=memory thing and then disabled it. I just now turned it on again.
mailbox_list_index_disable: no
These aren't actually always disabled since the code is buggy..
I tried searching for what this did or what it meant but didnt find anything helpful. Should I set manually set this to yes or leave it?
BTW. Are you using Dovecot's deliver for new mails?
No I am just using exim 4.69 to deliver directly to maildir