shrinking memory consumptions

[Jim Gettys]

The most we can hope for is that this happens *most* of the time.  When
the kernel needs memory in extremis, it will have to do what it has to
do.  The problem is a process can cause memory allocation so fast that
nothing else gets a chance to get in edgewise.

As I've said, the Nokia folks have been working in this area.  I hope to
talk to them at FISL (many of them are Brazilian).
                               - Jim


On Mon, 2007-04-02 at 14:48 -0400, Ivan Krstić wrote:
> Jim Gettys wrote:
> > Oh, I understand all right.  And ran it by those who really understand
> > the Linux kernel.  You end up in all sorts of deadlock hell if you try
> > to rely on user space for anything; at most you can hint to user space
> > that memory is getting low.
> 
> No -- I explicitly said I'm proposing a non-polling, non-blocking
> notification mechanism that's _orthogonal_ to the existing OOM killer
> functionality. There's no deadlock hell here. You're calling the same
> thing a 'hint'. We're in violent agreement.
> 
> Fundamentally, this is a user experience issue. Applications
> disappearing without any explanation, and for reasons completely
> unrelated to user actions (such as running low on memory) clearly isn't
> the user experience we want to provide. A userland application quit
> chooser, driven by a low-memory hint from the kernel and with a sane
> default application preselected for quitting -- chosen by some of the
> heuristics you mention -- is pretty obviously superior from a user
> experience point of view. And if the user takes too long to answer or
> the remaining memory dries up because something is allocating it very
> quickly, well, let the OOM killer go to town.
> 
-- 
Jim Gettys
One Laptop Per Child