Subject: Re: 128Mbyte memory Pentium Pro problem
To: Jason Thorpe <thorpej@nas.nasa.gov>
From: Chris G Demetriou <Chris_G_Demetriou@ux2.sp.cs.cmu.edu>
List: current-users
Date: 08/05/1996 05:58:53
> The reason that's not been increased in the the source tree is because it 
> leads to pages being wasted in unused pmap overhead if you don't requre 
> the additional address space, such as a low memory machine.  That's a 
> double-stinger - you have little memory, don't need the extra address 
> space, but the pmap would use up a bunch of memory so it could map pages 
> you don't have.  Ick :-)  The real solution, of course, is to fix the 
> pmap, but thusfar, noone has stepped up to do so.

A fairly good (but not "real" 8-) solution would be to put some simple
estimator into locore, that does a better job than the current
constant.

it could be as easy as "some base value + (constant * memory size)"...
However, it'd be better if it took other factors (e.g. expected or
compiled-in size of buffer pool, number of mbuf cluster and kernel
malloc pages, etc.) into account.

If somebody's trying to install NetBSD at a new site, i.e. they have
no other machine running on which to compile a kernel, saying "compile
a kernel" won't help.  saying "pull out some memory, compile a kernel,
then put them memory back" would help, but won't win many friends.


cgd