Thanks; these look interesting. We have a similar nas setup but we have 2 frontend dovecot servers connecting to it and store the indexes over nfs. We also have around 10 mail servers running deliver to try to keep the indexes on the nfs store up-to-date. Have you done any tests with the speed of multiple boxes each maintaining a local index of the mailbox? I suspect in this case keeping indexes on nfs would be the best bet but I don't have anything to substantiate that claim...
Also one thing to note with storing things on nfs is that there are a large number of broken kernels out there (they issue about 10* more nfs lookup requests than they should) - centos 5.1 had these issues iirc (though the centosplus kernel and centos 5.2 did fix it). Always give it a good test before you change the kernel on your server... I assume you are using nfs3; has anyone tried using heavily loaded nfs4 and seeing if any better performance can be achieved?
Another thing - I found that dovecot's pop3 implimentation is worse than courier's over nfs (wait state on our boxes is significantly increased). I still don't really understand why this is; I suspect it's probably due to to the indexes being created/updated though I thought these were meant to be discontinued after a while if it is just a simple login/fetch all operation. I only mention this because if you are offering pop then you should really do the same benchmarks for that.
Mark