>Just as a mini data point (and I don't know what your '10% improvement'
>is measured on, but I'm seeing about a 40% improvement between a
>5000/240 and a 5000/260 for a kernel build (make clean depend all, from
>around 55 minutes to around 32 minutes).  With both of these machines,
>everything but the root fs is NFS.

Doing a build on a 5000/150 (or 5000/50), better-tuned pmap code shows
a 10% walltime improvement over using the pmap code currently in the tree.

The difference between a 20-33MHz 3000, and a 50 Mhz r4000 or 60 MHz
r4400, is something else again :)