lkml.org 
[lkml]   [2012]   [Jun]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 0/5] Some vmevent fixes...
    On Fri, Jun 08, 2012 at 02:45:07AM -0700, Anton Vorontsov wrote:
    > On Fri, Jun 08, 2012 at 09:12:36AM +0000, leonid.moiseichuk@nokia.com wrote:
    > [...]
    > > > Exactly. I don't know why people think pushing vmevents to userspace is
    > > > going to fix any of the hard problems.
    > > >
    > > > Anton, Lenoid, do you see any fundamental issues from userspace point of
    > > > view with going forward what Minchan is proposing?
    > >
    > > That good proposal but I have to underline that userspace could be interested not only in memory consumption stressed cases (pressure, vm watermarks ON etc.)
    > > but quite relaxed as well e.g. 60% durty pages are consumed - let's do not restart some daemons. In very stressed conditions user-space might be already dead.
    >
    > Indeed. Minchan's proposal is good to get notified that VM is under
    > stress.
    >
    > But suppose some app allocates memory slowly (i.e. I scroll a large
    > page on my phone, and the page is rendered piece by piece). So, in
    > the end we're slowly but surely allocate a lot of memory. In that
    > case Minchan's method won't notice that it's actually time to close
    > some apps.

    I can't understand. Why can't the approach catch the situation?
    Let's think about it.

    There is 40M in CleanCache LRU which has easy-reclaimable pages and
    there is 10M free pages and 5M high watermark in system.

    Your application start to consume free pages very slowly.
    So when your application consumed 5M, VM start to reclaim. So far, it's okay
    because we have 40M easy-reclaimable pages. And low memory notifier can start
    to notify so your dameon can do some action to get free pages.

    I think it's not so late.

    sidenote:
    It seems I live in the complete opposite place because
    you guys always start discussion when I am about going out of office.
    Please understand my late response.
    Maybe I will come back after weekend. :)

    Thanks.

    >
    > Then suppose someone calls me, the "phone" application is now
    > starting, but since we're out of 'easy to reclaim' pages, it takes
    > forever for the app to load, VM is now under huge stress, and surely
    > we're *now* getting notified, but alas, it is too late. Call missed.
    >
    >
    > So, it's like measuring distance, velocity and acceleration. In
    > Android case, among other things, we're interested in distance too!
    > I.e. "how much exactly 'easy to reclaim' pages left", not only
    > "how fast we're getting out of 'easy to reclaim' pages".
    >
    > > Another interesting question which combination of VM page types could be recognized as interesting for tracking as Minchan correctly stated it depends from area.
    > > For me seems weights most likely will be -1, 0 or +1 to calculate resulting values and thesholds e.g. Active = {+1 * Active_Anon; +1 * Active_File}
    > > It will extend flexibility a lot.
    >
    > Exposing VM details to the userland? No good. :-)
    >
    > --
    > Anton Vorontsov
    > Email: cbouatmailru@gmail.com


    \
     
     \ /
      Last update: 2012-06-08 13:21    [W:4.181 / U:0.088 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site