Re: Memory statistics

["Matthias Clasen" <matthias clasen gmail com>]

On 4/20/06, Federico Mena Quintero <federico ximian com> wrote:
> On Tue, 2006-04-18 at 19:52 +0200, Benoît Dejean wrote:
>
> > import os
> > import gtop
> >
> > for pid in gtop.proclist(gtop.PROCLIST_KERN_PROC_UID, os.getuid()):
> >       gtop.proc_map(<pid>)
>
> /me self-slaps for reinventing the wheel :)
>
> Some quick numbers from a bit of grep/awk action on the results:
>
> - The private_dirty for libraries is under 14 MB on my box.  This is
> much smaller than the heaps for fat apps (i.e. hundreds of MB for
> Evolution/Mozilla).  So, reducing .data and friends is probably
> interesting only for embedded devices.
>
> - We have a lot of mappings that are exactly 8192 KB.  This is the
> default stack size for threads.  Most of those mappings are clean (few
> or none pages committed to memory), which means that threads have
> smaller requirements for their stack size than the default.  Still, this
> looks bad in top(1) and in pmap :)
>
> - All GTK+ apps which display images have about 700 KB of shared memory
> segments, used for MIT-SHM with the X server.  The segments are totally
> private_dirty.  Could we free the segments if they haven't been used in
> N seconds?  Could we only start using them until we have to transfer a
> big image to the X server?  Transferring icons is probably fast enough
> over the normal Unix domain socket.  For my session, the shm segments
> occupy about 6 MB in total --- again, probably more useful for embedded
> devices than for desktop boxes.
>

the issue of the xshm memory consumption is
tracked in bug 169544, and i believe maemo has
a patch to reduce it.

Matthias