On Thu, 28 Oct 2004 at 14:44:53 +0300 Denis Vlasenko wrote:
> Even if I add up size of every process, *counting libc shared pages
> once per process* (which will overestimate memory usage), I arrive at
> 23mb *total memory required by all processes*. How come kernel
> found 90mb to swap out? There is NOTHING to swap out except those
> 23mb!
>
> (Of course when oom_trigger was running, kernel first swapped out
> those 23mb and then started swapping out momery taken by oom_trigger
> itself, but when oom_trigger was killed, its RAM *and* swapspace
> should be deallocated. Thus I expected to see ~20 mb swap usage).
I am seeing this with Mozilla in an Xnest session. Even after I terminate
Mozilla + Xnest there is a huge amount of swapped-out memory (~100MB). This
is on a system with 320MB of memory. Since the problem goes away when I
leave X I had assumed it was an X bug (Fedora Core 1, not up-to-date) but
now I wonder... Kernel version is 2.6.9 + patches from L-K but problem is
the same in base 2.6.9.
--Chuck Ebbert 28-Oct-04 10:48:07
On Thursday 28 October 2004 21:04, Chuck Ebbert wrote:
> On Thu, 28 Oct 2004 at 14:44:53 +0300 Denis Vlasenko wrote:
> > Even if I add up size of every process, *counting libc shared pages
> > once per process* (which will overestimate memory usage), I arrive at
> > 23mb *total memory required by all processes*. How come kernel
> > found 90mb to swap out? There is NOTHING to swap out except those
> > 23mb!
> >
> > (Of course when oom_trigger was running, kernel first swapped out
> > those 23mb and then started swapping out momery taken by oom_trigger
> > itself, but when oom_trigger was killed, its RAM *and* swapspace
> > should be deallocated. Thus I expected to see ~20 mb swap usage).
>
>
> I am seeing this with Mozilla in an Xnest session. Even after I terminate
> Mozilla + Xnest there is a huge amount of swapped-out memory (~100MB). This
> is on a system with 320MB of memory. Since the problem goes away when I
> leave X I had assumed it was an X bug (Fedora Core 1, not up-to-date) but
> now I wonder... Kernel version is 2.6.9 + patches from L-K but problem is
> the same in base 2.6.9.
I have no additional data yet, but it looks similar to the "oom triggers too
early while there's plenty of cache to get rid of" misbehavior which
was frequently seen in 2.4.10 era.
However, "90mb in swap on a system with 30 mb total virtual space of all
processes combined" is something entirely new :)
--
vda