Re: [PATCH 4/4] memcg: prevent from reclaiming if there are per-cpucached charges

From: KAMEZAWA Hiroyuki
Date: Thu Jul 21 2011 - 20:04:15 EST


On Thu, 21 Jul 2011 14:30:12 +0200
Michal Hocko <mhocko@xxxxxxx> wrote:

> On Thu 21-07-11 19:54:11, KAMEZAWA Hiroyuki wrote:
> > On Thu, 21 Jul 2011 10:28:10 +0200
> > Michal Hocko <mhocko@xxxxxxx> wrote:
> >
> > > If we fail to charge an allocation for a cgroup we usually have to fall
> > > back into direct reclaim (mem_cgroup_hierarchical_reclaim).
> > > The charging code, however, currently doesn't care about per-cpu charge
> > > caches which might have up to (nr_cpus - 1) * CHARGE_BATCH pre charged
> > > pages (the current cache is already drained, otherwise we wouldn't get
> > > to mem_cgroup_do_charge).
> > > That can be quite a lot on boxes with big amounts of CPUs so we can end
> > > up reclaiming even though there are charges that could be used. This
> > > will typically happen in a multi-threaded applications pined to many CPUs
> > > which allocates memory heavily.
> > >
> >
> > Do you have example and score, numbers on your test ?
>
> As I said, I haven't seen anything that would affect visibly performance
> but I have seen situations where we reclaimed even though there were
> pre-charges on other CPUs.
>
> > > Currently we are draining caches during reclaim
> > > (mem_cgroup_hierarchical_reclaim) but this can be already late as we
> > > could have already reclaimed from other groups in the hierarchy.
> > >
> > > The solution for this would be to synchronously drain charges early when
> > > we fail to charge and retry the charge once more.
> > > I think it still makes sense to keep async draining in the reclaim path
> > > as it is used from other code paths as well (e.g. limit resize). It will
> > > not do any work if we drained previously anyway.
> > >
> > > Signed-off-by: Michal Hocko <mhocko@xxxxxxx>
> >
> > I don't like this solution, at all.
> >
> > Assume 2 cpu SMP, (a special case), and 2 applications running under
> > a memcg.
> >
> > - one is running in SCHED_FIFO.
> > - another is running into mem_cgroup_do_charge() and call drain_all_stock_sync().
> >
> > Then, the application stops until SCHED_FIFO application release the cpu.
>
> It would have to back off during reclaim anyaway (because we check
> cond_resched during reclaim), right?
>

just have cond_resched() on a cpu which calls some reclaim stuff. It will no help.


> > In general, I don't think waiting for schedule_work() against multiple cpus
> > is not quicker than short memory reclaim.
>
> You are right, but if you consider small groups then the reclaim can
> make the situation much worse.
>

If the system has many memory and the container has many cgroup, memory is not
small because ...to use cpu properly, you need memroy. It's a mis-configuration.



> > Adding flush_work() here means that a context switch is requred before
> > calling direct reclaim.
>
> Is that really a problem? We would context switch during reclaim if
> there is something else that wants CPU anyway.
> Maybe we could drain only if we get a reasonable number of pages back?
> This would require two passes over per-cpu caches to find the number -
> not nice. Or we could drain only those caches that have at least some
> threshold of pages.
>
> > That's bad. (At leaset, please check __GFP_NOWAIT.)
>
> Definitely a good idea. Fixed.
>
> > Please find another way, I think calling synchronous drain here is overkill.
> > There are not important file caches in the most case and reclaim is quick.
>
> This is, however, really hard to know in advance. If there are used-once
> unmaped file pages then it is much easier to reclaim them for sure.
> Maybe I could check the statistics and decide whether to drain according
> pages we have in the group. Let me think about that.
>
> > (And async draining runs.)
> >
> > How about automatically adjusting CHARGE_BATCH and make it small when the
> > system is near to limit ?
>
> Hmm, we are already bypassing batching if we are close to the limit,
> aren't we? If we get to the reclaim we fallback to nr_pages allocation
> and so we do not refill the stock.
> Maybe we could check how much we have reclaimed and update the batch
> size accordingly.
>

Please wait until "background reclaim" stuff. I don't stop it and it will
make this cpu-caching stuff better because we can drain before hitting
limit.

If you cannot wait....

One idea is to have a threshold to call async "drain". For example,

threshould = limit_of_memory - nr_online_cpu() * (BATCH_SIZE + 1)

if (usage > threshould)
drain_all_stock_async().

Then, situation will be much better.



Thanks,
-Kame



--
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/