>>And why do things become so slow when the list length grows to 150+ entries?

>i don't have a clear understanding of the slowdown, so it is hard to say.
>you could try a uvm kernel and see if that makes any difference.

John Dyson's observation is that the Mach-derived VM code doesn't
actually merge address-space descriptors on the freelist. Allocating
and freeing chunks of `adjacent' memory can result in long linked
lists inside the kernel. 

One hypothesis is that the slowdown is due to the mmap()/munmap()
calls by XFree86's InternalMalloc, which results in long VM lists and
frequent traversals of them during certain client request streams to
the server.

If that's correct, and if UVM merges fragments of VM space on
munmap(), maybe the slowdown will go away.