>In the dynamic linked case, some libary function is probably mapped
>at an address that uses the same cache line(s) as some fairly heavily
>used loop or variable in the compuation. ...

This is possible (and was my first suspicion), but then the code would
run equally fast (slow) single and multi user.

Incidentally, the above behavior is why so many machines use either
multiway set-associative caches or (the new `hot' idea) `victim caches'.
(See comp.arch.)

Chris