Re: [PATCH -mmotm 0/5] memcg: per cgroup dirty limit (v6)

From: Andrea Righi
Date: Thu Mar 11 2010 - 17:23:58 EST


On Thu, Mar 11, 2010 at 09:39:13AM +0900, KAMEZAWA Hiroyuki wrote:
> On Wed, 10 Mar 2010 00:00:31 +0100
> Andrea Righi <arighi@xxxxxxxxxxx> wrote:
>
> > Control the maximum amount of dirty pages a cgroup can have at any given time.
> >
> > Per cgroup dirty limit is like fixing the max amount of dirty (hard to reclaim)
> > page cache used by any cgroup. So, in case of multiple cgroup writers, they
> > will not be able to consume more than their designated share of dirty pages and
> > will be forced to perform write-out if they cross that limit.
> >
> > The overall design is the following:
> >
> > - account dirty pages per cgroup
> > - limit the number of dirty pages via memory.dirty_ratio / memory.dirty_bytes
> > and memory.dirty_background_ratio / memory.dirty_background_bytes in
> > cgroupfs
> > - start to write-out (background or actively) when the cgroup limits are
> > exceeded
> >
> > This feature is supposed to be strictly connected to any underlying IO
> > controller implementation, so we can stop increasing dirty pages in VM layer
> > and enforce a write-out before any cgroup will consume the global amount of
> > dirty pages defined by the /proc/sys/vm/dirty_ratio|dirty_bytes and
> > /proc/sys/vm/dirty_background_ratio|dirty_background_bytes limits.
> >
> > Changelog (v5 -> v6)
> > ~~~~~~~~~~~~~~~~~~~~~~
> > * always disable/enable IRQs at lock/unlock_page_cgroup(): this allows to drop
> > the previous complicated locking scheme in favor of a simpler locking, even
> > if this obviously adds some overhead (see results below)
> > * drop FUSE and NILFS2 dirty pages accounting for now (this depends on
> > charging bounce pages per cgroup)
> >
> > Results
> > ~~~~~~~
> > I ran some tests using a kernel build (2.6.33 x86_64_defconfig) on a
> > Intel Core 2 @ 1.2GHz as testcase using different kernels:
> > - mmotm "vanilla"
> > - mmotm with cgroup-dirty-memory using the previous "complex" locking scheme
> > (my previous patchset + the fixes reported by Kame-san and Daisuke-san)
> > - mmotm with cgroup-dirty-memory using the simple locking scheme
> > (lock_page_cgroup() with IRQs disabled)
> >
> > Following the results:
> > <before>
> > - mmotm "vanilla", root cgroup: 11m51.983s
> > - mmotm "vanilla", child cgroup: 11m56.596s
> >
> > <after>
> > - mmotm, "complex" locking scheme, root cgroup: 11m53.037s
> > - mmotm, "complex" locking scheme, child cgroup: 11m57.896s
> >
> > - mmotm, lock_page_cgroup+irq_disabled, root cgroup: 12m5.499s
> > - mmotm, lock_page_cgroup+irq_disabled, child cgroup: 12m9.920s
> >
> > With the "complex" locking solution, the overhead introduced by the
> > cgroup dirty memory accounting is minimal (0.14%), compared with the overhead
> > introduced by the lock_page_cgroup+irq_disabled solution (1.90%).
> >
> Hmm....isn't this bigger than expected ?

Consider that I'm not running the kernel build on tmpfs, but on a fs
defined on /dev/sda. So the additional overhead should be normal
compared to the mmotm vanilla, where there's only FILE_MAPPED
accounting.

-Andrea
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/