the "RSS" value is not counting pages in memory, it's counting the number
of pmap entries for the process, which is only an approximation of the
number of pages that the process cares about which might in memory.
it's probably a pretty good approximation, though.

if this is all anonymous memory allocated with malloc(), you can adjust
the sysctl knob for "vm.anonmin" to be more like 75 instead of the
default value of 10.  I don't know what the total effect of this will
on your machine, though.

-Chuck


On Fri, Nov 09, 2001 at 04:28:04PM -0800, Michael Graff wrote:
> I have several machines running -current, and I'm seeing some rather
> annoying memory behavior.
> 
> I have a process that takes about 700 MB, but on a 1 G machine very
> little of it is actually allowed to be resident at any point:
> 
> PID USERNAME PRI NICE   SIZE   RES STATE      TIME   WCPU    CPU   COMMAND
> 340 explorer  -5    4   712M  108M biowai/1  37:01 33.59% 33.59%   foo
> 
> ps axl returns the same data, so it's not a "top" problem.
> 
> On a 1.5 G system, approximately the same thing happens.  It's as if
> there is an internal limit on the size a process can grow to before
> the buffer cache becomes more important.
> 
> It does burst up to about 210 MB or so resident, then seems to page
> back down again to 100 MB or so.
> 
> Are there knobs I can tweak here?  I'm certainly not a VM expert, but
> I'm going to poke around a bit, too.
> 
> --Michael
> 
> 
>