"Stephen M. Jones" <smj@cirr.com> writes:

> directory.  When I do this with two CS20s running FreeBSD 5.2.1, I do not
> generate a load over 1.00 and I do not experience ethernet timeouts.

You don't actually say so, but does this imply that you do see
timeouts on NetBSD?

> 
> However, within a few hours of doing the same excerise using a NetBSD 2.0F
> client writing to the NFS mounted filesystem I have a process table that
> looks like this:

[four [nfsio]'s with about 3:40 time each, load 4.8]

> The CS20 running FreeBSD 5.2.1 is SMP however and has a proc table that looks
> like this:

[five nfsd's with more varying time, load 0.02]

> I can probably kill off the nfsio processes and bring the load down, but
> my question is, why is it happening?  When I have the scenario reversed
> you might find it interesting that FreeBSD as a client has delays (nfsrcvlk)
> during simple operations such as 'cat' and 'ls' while the 500mb file is 
> being created.  When NetBSD is the client 'cat' and 'ls' are pretty snappy.

Is the load number a problem? Is the system otherwise less responsive?
I think you're seeing artifacts of how load is measured on the two
systems, and a more round-robin distribution of NFS tasks to kernel
threads in the case of NetBSD. In other words, a somewhat
apples-to-oranges comparison. You need to find a benchmark result you
care about, and measure that.

        - Nathan