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

[Michal Hocko]

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? 

> 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.

> 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.

> or flushing ->stock periodically ?
> 
> 
> Thanks,
> -Kame

Thanks!
-- 
Michal Hocko
SUSE Labs
SUSE LINUX s.r.o.
Lihovarska 1060/12
190 00 Praha 9    
Czech Republic
--
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/