On Mon, Jul 16, 2001 at 01:43:34PM -0500, Dimitri Maziuk wrote: | * D-Man ([EMAIL PROTECTED]) spake thusly: | ... | > | > Now suppose just the right packets are lost and the RPC call ends up | > matching a different, existant, procedure that doesn't have the | > intended effect <grin> ... sounds like it would be a good idea to | > make NFS over TCP stable :-). | | Well, RPC has its own error correction so if a right packet is | lost it will be re-transmitted. Or RPC call will time out and | return error (d'oh! remote RPC call. Automated ATM machine). | The difference is that this is handled by application-level (if | you consider RPC to be in application layer) code, not by transport | layer. | | I imagine implementing NFS over TCP would involve a re-design of | RPC state machine and a serious re-write of all related code, and | it ain't exactly broken as it is, so... (given all the things that | could [theoretically] go wrong with NFS, it is surprisingly stable).
Hmm, yeah, I guess that could be hard, unless the RPC mechanism could use TCP. Or maybe a different RPC implementation could be used that would work over TCP. Or maybe it isn't really a problem in practice but just in theory. | > Can I use NFS-root-over-TCP for one of the boxes (I'll have 2, the | > other I'll leave at "regular" UDP as a "control" system)? | | There are other networked file systems out there, like Coda, more modern | and arguably better than NFS. If you only need to support linux, why not | use one of them? Or [e]nbd? They're both going to be Debian. Can I use root-over-$FS for those filesystems? Are they stable? -D