Re: TCP-based NFS?

Craig Milo Rogers (rogers@isi.edu)
Sat, 07 Sep 96 12:39:09 PDT


>Hmm, well TCP has always been a can do this option for NFS implementers
>AFAIK, but most don't support it, but I think on an uncongested ethernet the
>TCP over head is much greater, hence UDP.

WHen NFS was first implement, TCP was much slower than UDP.
This is not, however, true in general. TCP is much more complex to
implement correctly, but on a well-tuned implementation the overhead
is not significantly greater (on a modern processor, using a 10-Mbps
LAN).

There is one way in which UDP's overhead is lower than TCP's:
data checksums are optional in UDP (in IPv4). But:

1) NFS over UDP without checksums is a <b>bad idea</b>. People
have lost filesystems because of it.

2) On many (most?) modern computers, the checksum calculation can be
combined with a buffer copy at no additional cost.

Craig Milo Rogers