Re: [Devel] Re: [RFC][PATCH 2/7] RSS controller core

From: Andrew Morton
Date: Tue Mar 13 2007 - 06:50:23 EST


> On Tue, 13 Mar 2007 13:19:53 +0300 Kirill Korotaev <dev@xxxxx> wrote:
> Andrew Morton wrote:
> >>>> - shared mappings of 'shared' files (binaries
> >>>> and libraries) to allow for reduced memory
> >>>> footprint when N identical guests are running
> >>>
> >>>So, it sounds like this can be phrased as a requirement like:
> >>>
> >>> "Guests must be able to share pages."
> >>>
> >>>Can you give us an idea why this is so?
> >>
> >>sure, one reason for this is that guests tend to
> >>be similar (or almost identical) which results
> >>in quite a lot of 'shared' libraries and executables
> >>which would otherwise get cached for each guest and
> >>would also be mapped for each guest separately
> >
> >
> > nooooooo. What you're saying there amounts to text replication. There is
> > no proposal here to create duplicated copies of pagecache pages: the VM
> > just doesn't support that (Nick has soe protopatches which do this as a
> > possible NUMA optimisation).
> >
> > So these mmapped pages will contiue to be shared across all guests. The
> > problem boils down to "which guest(s) get charged for each shared page".
> >
> > A simple and obvious and easy-to-implement answer is "the guest which paged
> > it in". I think we should firstly explain why that is insufficient.
> I guess by "paged it in" you essentially mean
> "mapped the page into address space for the *first* time"?

Not really - I mean "first allocated the page". ie: major fault(), read(),
write(), etc.

> i.e. how many times the same page mapped into 2 address spaces
> in the same container should be accounted for?
>
> We believe ONE. It is better due to:
> - it allows better estimate how much RAM container uses.
> - if one container mapped a single page 10,000 times,
> it doesn't mean it is worse than a container which mapped only 200 pages
> and that it should be killed in case of OOM.

I'm not sure that we need to account for pages at all, nor care about rss.

If we use a physical zone-based containment scheme: fake-numa,
variable-sized zones, etc then it all becomes moot. You set up a container
which has 1.5GB of physial memory then toss processes into it. As that
process set increases in size it will toss out stray pages which shouldn't
be there, then it will start reclaiming and swapping out its own pages and
eventually it'll get an oom-killing.

No RSS acounting or page acounting in sight, because we already *have* that
stuff, at the physical level, in the zone.

Overcommitment can be performed by allowing different containers to share
the same zone set, or by dynamically increasing or decreasing the size of
a physical container.

This all works today with fake-numa and cpusets, no kernel changes needed.

It could be made to work fairly simply with a multi-zone approach, or with
resizeable zones.

I'd be interested in knowing what you think the shortcomings of this are
likely to be,.

-
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/