Re: [RFC][PATCH 0/4] Memory controller soft limit patches (v2)

[Balbir Singh]

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2009-02-17 14:10:39]:

> On Tue, 17 Feb 2009 10:11:10 +0530
> Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> 
> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2009-02-17 13:03:52]:
> > 
> > > On Tue, 17 Feb 2009 08:35:26 +0530
> > > Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> > > I don't want to add any new big burden to kernel hackers of memory management,
> > > they work hard to improve memory reclaim. This patch will change the behavior.
> > > 
> > 
> > I don't think I agree, this approach suggests that before doing global
> > reclaim, there are several groups that are using more than their
> > share of memory, so it makes sense to reclaim from them first.
> > 
> 
> > 
> > > BTW, in typical bad case, several threads on cpus goes into memory recalim at once and
> > > all thread will visit this memcg's soft-limit tree at once and soft-limit will
> > > not work as desired anyway.
> > > You can't avoid this problem at alloc_page() hot-path.
> > 
> > Even if all threads go into soft-reclaim at once, the tree will become
> > empty after a point and we will just return saying there are no more
> > memcg's to reclaim from (we remove the memcg from the tree when
> > reclaiming), then those threads will go into regular reclaim if there
> > is still memory pressure.
> 
> Yes. the largest-excess group will be removed. So, it seems that it doesn't work
> as designed. rbtree is considered as just a hint ? If so, rbtree seems to be
> overkill.
> 
> just a question:
> Assume memcg under hierarchy.
>    ../group_A/                 usage=1G, soft_limit=900M  hierarchy=1
>               01/              usage=200M, soft_limit=100M
>               02/              usage=300M, soft_limit=200M
>               03/              usage=500M, soft_limit=300M  <==== 200M over.
>                  004/          usage=200M, soft_limit=100M
>                  005/          usage=300M, soft_limit=200M
> 
> At memory shortage, group 03's memory will be reclaimed 
>   - reclaim memory from 03, 03/004, 03/005
> 
> When 100M of group 03' memory is reclaimed, group_A 's memory is reclaimd at the
> same time, implicitly. Doesn't this break your rb-tree ?
> 
> I recommend you that soft-limit can be only applied to the node which is top of
> hierarchy.

Yes, that can be done, but the reason for putting both was to target
the right memcg early.

> 
> > > 
> > > > > 3. After this patch, res_counter is no longer for general purpose res_counter...
> > > > >    It seems to have too many unnecessary accessories for general purpose.  
> > > > 
> > > > Why not? Soft limits are a feature of any controller. The return of
> > > > highest ancestor might be the only policy we impose right now. But as
> > > > new controllers start using res_counter, we can clearly add a policy
> > > > callback.
> > > > 
> > > I think you forget that memroy cgroups is an only controller in which the kernel
> > > can reduce the usage of resource without any harmful to users.
> > > soft-limit is nonsense for general resources, I think.
> > > 
> > 
> > Really? Even for CPUs? soft-limit is a form of shares (please don't
> > confuse with cpu.shares). Soft limits is used as a way of implementing
> > work conserving controllers.
> > 
> 
> I don't think cpu needs this. It works under share and no hardlimit.
>

Forget CPUs for now. The concept of soft-limits is applicable to all
resource controllers. Look at check_thread_timers, you'll see CPU soft
limits for rlimit. soft limits allow overcommit as long as there is no
contention on the resource. 

-- 
	Balbir
--
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/