Does this mean you had the directory containing the
   kernels NFS exported too? As far as I know, the alpha bootblocks
   can't TFTP the kernel.

Yes, we actually have several Alphas running NetBSD, and use the largest
as a development machine.  We nfs-export several directories on that for
the kernels, so each client can get a kernel just built on the development
machine for testing the devices on that particular client.  If we like the
results, we copy the kernel over to the disk farm and put it on the nfs-
exported root file system.  This isn't required, just pleasant in our
circumstance.
   
   > perhaps can we add new variables
   > that the kernel will use to support separate tftp and nfs servers
   > (some sort of remote nfs path, and an nfs server field)? 
   
   I'm a bit confused - as things are now, only the console
   does TFTP for the "netboot" program.
   Do you want separate "kernel NFS" and "root NFS"
   servers? If yes, a workaround would be to use the old
   bootparam in either the bootblocks or the kernel. Both
   should be easily done - a cpp definition for the bootblocks or
   a variable (nfs_boot_rfc951) in the kernel.
   If you want to TFTP the kernel, you can try to
   pull some code from the i386 netboot.
   And if you really only want to separate TFTP and
   NFS servers, you can use the "swap server" BOOTP
   tag - it will be used as NFS server if present.
   
At the moment, the 1.3 alpha release seems to use the "sa" value as
the tftp server, and the bootp server address as the nfs server.  I
was surprised by this, too, perhaps it is just the message from the
kernel saying "root is on ..." that is wrong, I haven't investigated
the source code yet.  I'll try the swap server tag and see what happens,
thanks for the suggestion.  (I don't want to tftp the kernel. Just have
a choice of kernels, and we did that by server fs division.  I think cjs
understood the wants in his response.)