Re: [patch 0/8] mm: memcg naturalization -rc2

[Ying Han]

On Thu, Jun 9, 2011 at 11:36 AM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> On Thu, Jun 09, 2011 at 10:36:47AM -0700, Ying Han wrote:
>> On Thu, Jun 9, 2011 at 1:35 AM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>> > On Wed, Jun 08, 2011 at 08:52:03PM -0700, Ying Han wrote:
>> >> On Wed, Jun 8, 2011 at 8:32 AM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
>> >> > I guess it would make much more sense to evaluate if reclaiming from
>> >> > memcgs while there are others exceeding their soft limit is even a
>> >> > problem.  Otherwise this discussion is pretty pointless.
>> >>
>> >> AFAIK it is a problem since it changes the spec of kernel API
>> >> memory.soft_limit_in_bytes. That value is set per-memcg which all the
>> >> pages allocated above that are best effort and targeted to reclaim
>> >> prior to others.
>> >
>> > That's not really true.  Quoting the documentation:
>> >
>> >    When the system detects memory contention or low memory, control groups
>> >    are pushed back to their soft limits. If the soft limit of each control
>> >    group is very high, they are pushed back as much as possible to make
>> >    sure that one control group does not starve the others of memory.
>> >
>> > I am language lawyering here, but I don't think it says it won't touch
>> > other memcgs at all while there are memcgs exceeding their soft limit.
>>
>> Well... :) I would say that the documentation of soft_limit needs lots
>> of work especially after lots of discussions we have after the LSF.
>>
>> The RFC i sent after our discussion has the following documentation,
>> and I only cut & paste the content relevant to our conversation here:
>>
>> What is "soft_limit"?
>> The "soft_limit was introduced in memcg to support over-committing the
>> memory resource on the host. Each cgroup can be configured with
>> "hard_limit", where it will be throttled or OOM killed by going over
>> the limit. However, the allocation can go above the "soft_limit" as
>> long as there is no memory contention. The "soft_limit" is the kernel
>> mechanism for re-distributing spare memory resource among cgroups.
>>
>> What we have now?
>> The current implementation of softlimit is based on per-zone RB tree,
>> where only the cgroup exceeds the soft_limit the most being selected
>> for reclaim.
>>
>> It makes less sense to only reclaim from one cgroup rather than
>> reclaiming all cgroups based on calculated propotion. This is required
>> for fairness.
>>
>> Proposed design:
>> round-robin across the cgroups where they have memory allocated on the
>> zone and also exceed the softlimit configured.
>>
>> there was a question on how to do zone balancing w/o global LRU. This
>> could be solved by building another cgroup list per-zone, where we
>> also link cgroups under their soft_limit. We won't scan the list
>> unless the first list being exhausted and
>> the free pages is still under the high_wmark.
>>
>> Since the per-zone memcg list design is being replaced by your
>> patchset, some of the details doesn't apply. But the concept still
>> remains where we would like to scan some memcgs first (above
>> soft_limit) .
>
> I think the most important thing we wanted was to round-robin scan all
> soft limit excessors instead of just the biggest one.  I understood
> this is the biggest fault with soft limits right now.
>
> We came up with maintaining a list of excessors, rather than a tree,
> and from this particular implementation followed naturally that this
> list is scanned BEFORE we look at other memcgs at all.
>
> This is a nice to have, but it was never the primary problem with the
> soft limit implementation, as far as I understood.
>
>> > It would be a lie about the current code in the first place, which
>> > does soft limit reclaim and then regular reclaim, no matter the
>> > outcome of the soft limit reclaim cycle.  It will go for the soft
>> > limit first, but after an allocation under pressure the VM is likely
>> > to have reclaimed from other memcgs as well.
>> >
>> > I saw your patch to fix that and break out of reclaim if soft limit
>> > reclaim did enough.  But this fix is not much newer than my changes.
>>
>> My soft_limit patch was developed in parallel with your patchset, and
>> most of that wouldn't apply here.
>> Is that what you are referring to?
>
> No, I meant that the current behaviour is old and we are only changing
> it only now, so we are not really breaking backward compatibility.
>
>> > The second part of this is:
>> >
>> >    Please note that soft limits is a best effort feature, it comes with
>> >    no guarantees, but it does its best to make sure that when memory is
>> >    heavily contended for, memory is allocated based on the soft limit
>> >    hints/setup. Currently soft limit based reclaim is setup such that
>> >    it gets invoked from balance_pgdat (kswapd).
>>
>> We had patch merged which add the soft_limit reclaim also in the global ttfp.
>>
>> memcg-add-the-soft_limit-reclaim-in-global-direct-reclaim.patch
>>
>> > It's not the pages-over-soft-limit that are best effort.  It says that
>> > it tries its best to take soft limits into account while reclaiming.
>> Hmm. Both cases are true. The best effort pages I referring to means
>> "the page above the soft_limit are targeted to reclaim first under
>> memory contention"
>
> I really don't know where you are taking this from.  That is neither
> documented anywhere, nor is it the current behaviour.
>
> Yeah, currently the soft limit reclaim cycle preceeds the generic
> reclaim cycle.  But the end result is that other memcgs are reclaimed
> from as well in both cases.  The exact timing is irrelevant.
>
> And this has been the case for a long time, so I don't think my rework
> breaks existing users in that regard.
>
>> > My code does that, so I don't think we are breaking any promises
>> > currently made in the documentation.
>> >
>> > But much more important than keeping documentation promises is not to
>> > break actual users.  So if you are yourself a user of soft limits,
>> > test the new code pretty please and complain if it breaks your setup!
>>
>> Yes, I've been running tests on your patchset, but not getting into
>> specific configurations yet. But I don't think it is hard to generate
>> the following scenario:
>>
>> on 32G machine, under root I have three cgroups with 20G hard_limit and
>> cgroup-A: soft_limit 1g, usage 20g with clean file pages
>> cgroup-B: soft_limit 10g, usage 5g with clean file pages
>> cgroup-C: soft_limit 10g, usage 5g with clean file pages
>>
>> I would assume reclaiming from cgroup-A should be sufficient under
>> global memory pressure, and no pages needs to be reclaimed from B or
>> C, especially both of them have memory usage under their soft_limit.
>
> Keep in mind that memcgs are scanned proportionally to their size,
> that we start out with relatively low scan counts, and that the
> priority levels are a logarithmic scale.
>
> The formula is essentially this:
>
>        (usage / PAGE_SIZE) >> priority
>
> which means that we would scan as follows, with decreased soft limit
> priority for A:
>
>        A: ((20 << 30) >> 12) >> 11 = 2560 pages
>        B: (( 5 << 30) >> 12) >> 12 =  320 pages
>        C:                          =  320 pages.
>
> So even if B and C are scanned, they are only shrunk by a bit over a
> megabyte tops.  For decreasing levels (if they are reached at all if
> there is clean cache around):
>
>        A: 20M 40M 80M 160M ...
>        B:  2M  4M  8M  16M ...
>
> While it would be sufficient to reclaim only from A, actually
> reclaiming from B and C is not a big deal in practice, I would
> suspect.

One way to think about how the user will set the soft_limit ( in our
case for example) is to set the soft_limit to be the working_set_size
of the cgroup (by doing working set estimation).

The soft_limit will be readjust at run time based on the workload. In
that case, shrinking the memory from B/C has potential performance
impact on the application. While it doesn't mean we can never reclaim
pages from them, but shrinking from A ( usage is 19G above its
soft_limit ) first will provide better predictability of performance.

--Ying

>
--
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/