On Tue, Apr 03, 2001 at 06:08:35PM +0200, Frank van der Linden wrote:
> On Tue, Apr 03, 2001 at 08:55:51AM -0700, Chuck Silvers wrote:
> > I'll question your math here...
> > looking at the netscape on my machine, the total text size for netscape
> > and all the shared libraries it uses is about 14 MB, and most data-section
> > pages are probably modified and thus would could as "anon" pages in UVM
> > and not "vtext".  this should easily fit in 25 MB (if shared libs were
> > accounted properly, which they're not).
> 
> On my machine:
> 
>   894 fvdl       2    0    20M   15M run       0:01  3.04%  2.83% netscape
>   238 fvdl       2    0  6248K   53M sleep     1:09  1.22%  1.22% XF86_SVGA

all of the memory usage numbers here include both cached executable files
("vtext") and process anonymous data ("anon").  these have separate limits.
vtext usage is composed of the text + data section sizes in the executable
(really just the text section), so looking as ps output isn't going to be
very informative.


> ..and this while netscape was only just started up. I do believe that
> this can exceed the 25M. I had to try hard, but I did manage to get
> netscape partly paged out by running an 1G bonnie and not touching
> it for a bit (on a 512M machine). And imagine a 128M machine, on
> which 5% is some 6M.

6 MB is pretty small, yea.  we probably do need to do something better
for smaller memory machines.  I'd say we should probably say we should
use 64 MB as the design point for a "small-memory workstation"; anything
smaller than that just isn't going to run netscape well no matter what
we do.  or is someone going to argue that they live comfortably in 32 MB?

whatever memory size we pick, we can run some sample workloads with
physical memory artificially limited to that size and see how various
tunings and code changes affect the interactive response.

-Chuck