On Tue, Aug 21, 2012 at 2:29 AM, Glauber Costa <glommer@xxxxxxxxxxxxx> wrote:
On 08/20/2012 10:30 PM, Ying Han wrote:Keep in mind that not all environments will be specialized to the point
> Not exactly. Here reclaiming from root is mainly for "reclaiming from
> root's exclusive lru", which links the page includes:
> 1. processes running under root
> 2. reparented pages from rmdir memcg under root
> 3. bypassed pages
>
> Setting root cgroup's softlimit = 0 has the implication of putting
> those pages to likely to reclaim, which works fine. The question is
> that if no other memcg is above its softlimit, would it be a problem
> to adding a bit extra pressure to root which always is eligible for
> softlimit reclaim ( usage is always greater than softlimit).
>
> As an example, it works fine in our environment since we don't
> explicitly put any process under root. Most of the pages linked in
> root lru would be reparented pages which should be reclaimed prior to
> others.
of having root memcg empty. This basically treats root memcg as a trash
bin, and can be very detrimental to use cases where actual memory is
present in there.
It would maybe be better to have all this garbage to go to a separate
place, like a shadow garbage memcg, which is invisible to the
filesystem, and is always the first to be reclaimed from, in any
circumstance.
We can certainly do something like that, and actually we have the *special* cgroup setup today in google's environment. It is mainly targeting for pages that are allocated not on behalf of applications, but more of
system maintainess overhead. One example would be kernel thread memory charging.
In this case, it might make sense to put those reparented pages to a separate cgroup. However I do wonder with the following questions:
1. it might only make sense to do that if something else running under root. As we know, root is kind of special in memcg where there is no limit on it. So I wonder what would be the real life use case to put something under root?
2. even the reparented pages are mixed together with pages from process running under root, the LRU mechanism should still take effect of evicting cold pages first. if the reparent pages are the left-over pages from the removed cgroups, I would assume they are the candidate to reclaim first.
I am curious that in your environment, do you have things running root?
--Ying