I'd be interested in hearing about this too. There are a number of
issues I know aren't right- but I just got through making a very
beefy kernel to manage 2GB of primary memory on an 8200- and all
I really had to do (for a *very* rough first approximation) was tweak
some parameters.

The current pmap code uses the page tables set up by the prom. It doesn't
use address space identfiers, and the amount of kernel heap space is
determined at pmap bootstrap time. This is all a problem, but it isn't
necessarily a showstopper for just being able to deal with a very fat
kernel, no?