Re: [RFI] Shared accounting for memory resource controller

[Balbir Singh]

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2009-04-07 16:00:14]:

> On Tue, 7 Apr 2009 12:07:22 +0530
> Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> 
> > Hi, All,
> > 
> > This is a request for input for the design of shared page accounting for
> > the memory resource controller, here is what I have so far
> > 
> 
> In my first impression, I think simple counting is impossible.
> IOW, "usage count" and "shared or not" is very different problem.
> 
> Assume a page and its page_cgroup.
> 
> Case 1)
>   1. a page is mapped by process-X under group-A
>   2. its mapped by process-Y in group-B (now, shared and charged under group-A)
>   3. move process-X to group-B
>   4. now the page is not shared.
> 

By shared I don't mean only between cgroups, it could be a page shared
in the same cgroup

> Case 2)
>   swap is an object which can be shared.
> 

Good point, I expect the user to account all cached pages as shared -
no?

> Case 3)
>   1. a page known as "A" is mapped by process-X under group-A.
>   2. its mapped by process-Y under group-B(now, shared and charged under group-A)
>   3. Do copy-on-write by process-X.
>      Now, "A" is mapped only by B but accoutned under group-A.
>      This case is ignored intentionally, now.

Yes, that is the original design

>      Do you want to call try_charge() both against group-A and group-B
>      under process-X's page fault ?
> 

No we don't, but copy-on-write is caught at page_rmap_dup() - no?

> There will be many many corner case.
> 
> 
> > Motivation for shared page accounting
> > -------------------------------------
> > 1. Memory cgroup administrators will benefit from the knowledge of how
> >    much of the data is shared, it helps size the groups correctly.
> > 2. We currently report only the pages brought in by the cgroup, knowledge
> >    of shared data will give a complete picture of the actual usage.
> > 
> 
> Motivation sounds good. But counting this in generic rmap will have tons of
> troubles and slow-down.
> 
> I bet we should prepare a file as
>   /proc/<pid>/cgroup_maps
> 
> And show RSS/RSS-owned-by-us per process. Maybe this feature will be able to be
> implemented in 3 days.

Yes, we can probably do that, but if we have too many processes in one
cgroup, we'll need to walk across all of them in user space. One other
alternative I did not mention is to walk the LRU like we walk page
tables and look at page_mapcount of every page, but that will be
very slow.

> 
> Thanks,
> -Kame
> 
> 

-- 
	Balbir
--
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/