I think you've nailed the problem with read cacheing against WfW shares.
Everything works fine now, with read speeds averaging 50MB/sec. for small
overall sizes <g>.
Now, back to the problem of nested NFS mounts...
My primary NFS server has two physical drives, with the secondary volume
mounted as /usr/src on the root. I have exported both of these in
/etc/exports, and verified that knfs picks them up into /var/lib/nfs/xtab.
The client machines can see the export using showmount -e. Mounting the
root directory of the server proceeds as expected. When I then attempt to
nest the /usr/src export in the client's mount hierarchy, things go badly
wrong.
The exact (mis)behavior depends upon whether the client is running the old
2.0.x or 2.1.x kernel. With older kernels, the first mount disappears
completely, and the second never shows up in the output from 'mount' (no
args). With newer kernels, only the / directory on the server vanishes!
The /usr/src directory (which, remember, is on the same physical volume as
/) _does_ appear, as does the /usr/src hierarchy. Sometimes (I can't
really repeat this with reliability), the /usr/src volume _appears_ to
mount correctly, but when you cd to the mount point you will find yourself
back at the base of the mount-point on the client!
If I mount / and /usr/src on unrelated points in the client filesystem, it
is just fine. The problem seems strictly related to nested mounts, which
should be legal and possible.
Let me know what you need for debugging, but I've got the feeling that
this is not going to be simple.
Steve