RFC: make the experimental NFS subsystem the default one

[Rick Macklem]

> >
> > Was it read rate you were referring to or something else?
> 
> Yes, read rate. Large file read performance was much better between
> two Solaris systems compared with Solaris as the server and FreeBSD as
> the client. This is when using gigabit ethernet.
> 
I know it sounds like I'm dodging the bullet, but my experience has been
that much slower has usually been related to the network layer. (There
were some krpc transport related issues in FreeBSD8 that could cause
abissmal performance is certain cases. I think/hope those are all fixed
now.)

Here I have a Solaris10 box where the net interface drops a packet repeatedly
under one load situation. I also have a laptop where the low end Realtek chip
drops packets, as reported by the stats from the chip (this latter one appears
to be a pure hardware issue). Both result in really slow NFS perf. when you
bump into them.

The traffic that NFS generates looks very different than what a bulk data
transfer's does. It tends to be small segments in both directions on the
TCP connection and then bursts of TCP segments going one way. (It is usually
a segment in the burst or a TCP segment going the opposite way at the
time of the burst that gets dropped.)

If you can easily reproduce the problem, you could capture a packet trace
(I wouldn't need a huge one, just enough of the slow stuff to see retransmits, etc)
and email it to me, if you want.

Others might be able to comment on good/not so good experiences with various
net chips/drivers they've used NFS with?

Anyhow you mentioned that you saw this on the regular NFS client, so I guess
I won't be in trouble if the new one is slow too:-)

rick