Re: [PATCH mmotm 2.5/4] memcg: disable irq at page cgroup lock (Re: [PATCH -mmotm 3/4] memcg: dirty pages accounting and limiting infrastructure)

[Andrea Righi <arighi@develer.com>]

On Thu, Mar 11, 2010 at 11:54:13AM -0500, Vivek Goyal wrote:
> On Thu, Mar 11, 2010 at 01:49:08PM +0900, KAMEZAWA Hiroyuki wrote:
> > On Thu, 11 Mar 2010 13:31:23 +0900
> > Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> wrote:
> > 
> > > On Wed, 10 Mar 2010 09:26:24 +0530, Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> > > > * nishimura@mxp.nes.nec.co.jp <nishimura@mxp.nes.nec.co.jp> [2010-03-10 10:43:09]:
> > 
> > > I made a patch(attached) using both local_irq_disable/enable and local_irq_save/restore.
> > > local_irq_save/restore is used only in mem_cgroup_update_file_mapped.
> > > 
> > > And I attached a histogram graph of 30 times kernel build in root cgroup for each.
> > > 
> > >   before_root: no irq operation(original)
> > >   after_root: local_irq_disable/enable for all
> > >   after2_root: local_irq_save/restore for all
> > >   after3_root: mixed version(attached)
> > > 
> > > hmm, there seems to be a tendency that before < after < after3 < after2 ?
> > > Should I replace save/restore version to mixed version ?
> > > 
> > 
> > IMHO, starting from after2_root version is the easist.
> > If there is a chance to call lock/unlock page_cgroup can be called in
> > interrupt context, we _have to_ disable IRQ, anyway.
> > And if we have to do this, I prefer migration_lock rather than this mixture.
> > 
> > BTW, how big your system is ? Balbir-san's concern is for bigger machines.
> > But I'm not sure this change is affecte by the size of machines.
> > I'm sorry I have no big machine, now.
> 
> FWIW, I took andrea's patches (local_irq_save/restore solution) and
> compiled the kernel on 32 cores hyperthreaded (64 cpus) with make -j32
> in /dev/shm/. On this system, I can't see much difference.
> 
> I compiled the kernel 10 times and took average.
> 
> Without andrea's patches: 28.698 (seconds)
> With andrea's patches: 28.711 (seconds).
> Diff is .04%
> 
> This is all should be in root cgroup. Note, I have not mounted memory cgroup
> controller but it is compiled in. So I am assuming that root group
> accounting will still be taking place. Also assuming that it is not
> required to do actual IO to disk and /dev/shm is enough to see the results
> of local_irq_save()/restore.

cgroup disable is at boot time "cgroup_disable=...", so root cgroup
accounting should be enabled.

The same for the local_irq_save/restore() overhead, lock/unlock_page_cgroup()
is called during each charge.

Many thanks for testing!
-Andrea