:-) I can see you saying "Oooooooooooh." (You would've had to been at dinner with Hitz tonight in Tokyo to understand the frame of mind I'm in.)
This is very interesting, actually.
You see, wrong or right I have this opinion that on clean, switched networks NFS/UDP seems fine (it is our default). NFS/TCP yields a 5-10% maximum throughtut hit (for things like SFS benchmark) and bigger whacks for sequential throughput benchmarks.
The argument is mostly made that for a WAN, the congestion control properties, and retransmit (in face of random packet loss) of TCP are preferred...
Per subnet specification would be cool (large groups of clients).
We are working in IETF NFS Version 4 workgroup on specs for next version. So I'm sort of interested in hearing new ideas.
beepy in Tokyo which is not snowing like it was in Korea on Thursday.
Oooooooooh. An idea is aborning....
How about allowing filers admins to specify the protocol capability profile that a client is allowed to see on a client-by-client basis in a file not unlike the name-mapping files and exports files and quotas files that we already use...
Then we could have NFS-over-TCP turned on for WAN clients but turned off for LAN clients and can experiment with NFS3/2 - all by configuring at the server end (one place) rather than at the client end (many places).
Thoughts?