lkml.org 
[lkml]   [2009]   [Apr]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFI] Shared accounting for memory resource controller
* KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-04-08 17:54:09]:

> On Wed, 8 Apr 2009 14:19:52 +0530
> Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
>
> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-04-08 17:03:41]:
> >
> > > On Wed, 8 Apr 2009 13:18:09 +0530
> > > Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> > >
> > > > > > 3. Using the above, we can then try to (using an algorithm you
> > > > > > proposed), try to do some work for figuring out the shared percentage.
> > > > > >
> > > > > This is the point. At last. Why "# of shared pages" is important ?
> > > > >
> > > >
> > > > I posted this in my motivation yesterday. # of shared pages can help
> > > > plan the system better and the size of the cgroup. A cgroup might have
> > > > small usage_in_bytes but large number of shared pages. We need a
> > > > metric that can help figure out the fair usage of the cgroup.
> > > >
> > > I don't fully understand but NR_FILE_MAPPED is an information in /proc/meminfo.
> > > I personally think I want to support information in /proc/meminfo per memcg.
> > >
> > > Hmm ? then, if you add a hook, it seems
> > > == mm/rmap.c
> > > 689 void page_add_file_rmap(struct page *page)
> > > 690 {
> > > 691 if (atomic_inc_and_test(&page->_mapcount))
> > > 692 __inc_zone_page_state(page, NR_FILE_MAPPED);
> > > 693 }
> > > == page_remove_rmap(struct page *page)
> > > 739 __dec_zone_page_state(page,
> > > 740 PageAnon(page) ? NR_ANON_PAGES : NR_FILE_MAPPED);
> > > ==
> > >
> > > Is good place to go, maybe.
> > >
> > > page->page_cgroup->mem_cgroup-> inc/dec counter ?
> > >
> > > Maybe the patch itself will be simple, overhead is unknown..
> >
> > I thought of the same thing, but then moved to the following
> >
> > ... mem_cgroup_charge_statistics(..) {
> > if (page_mapcount(page) == 0 && page_is_file_cache(page))
> > __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_RSS, val);
> >
> > But I've not yet tested the end result
> >
> I think
> - at uncharge:
> charge_statistics is only called when FILE CACHE is removed from radix-tree.
> mem_cgroup_uncharge() is called only when PageAnon(page).

Good point, I missed it, testing would have caught it.

> - at charge:
> charge_statistics is only called when FILE CACHE is added to radix-tree.
>
> This "checking only radix-tree insert/delete" help us to remove most of overheads
> on FILE CACHE.
>
> So, adding new hooks to page_add_file_rmap() and page_remove_rmap()
> is a way to go. (and easy to understand because we account it at the same time
> NR_FILE_MAPPED is modified.)

Agreed.

--
Balbir


\
 
 \ /
  Last update: 2009-04-08 11:05    [W:0.559 / U:0.160 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site