Re: [PATCH 4/4] Memory controller soft limit reclaim on contention (v6)

[Balbir Singh <balbir@linux.vnet.ibm.com>]

* KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-03-17 14:17:14]:

> On Tue, 17 Mar 2009 10:28:50 +0530
> Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> 
> > * KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> [2009-03-17 13:47:27]:
> > 
> > > On Tue, 17 Mar 2009 10:10:16 +0530
> > > Balbir Singh <balbir@linux.vnet.ibm.com> wrote:
> > > 
> > > > >   - vm.softlimit_ratio
> > > > > 
> > > > > If vm.softlimit_ratio = 99%, 
> > > > >   when sum of all usage of memcg is over 99% of system memory,
> > > > >   softlimit runs and reclaim memory until the whole usage will be below 99%.
> > > > >    (or some other trigger can be considered.)
> > > > > 
> > > > > Then,
> > > > >  - We don't have to take care of misc. complicated aspects of memory reclaiming
> > > > >    We reclaim memory based on our own logic, then, no influence to global LRU.
> > > > > 
> > > > > I think this approach will hide the all corner case and make merging softlimit 
> > > > > to mainline much easier. If you use this approach, RB-tree is the best one
> > > > > to go with (and we don't have to care zone's status.)
> > > > 
> > > > I like the idea in general, but I have concerns about
> > > > 
> > > > 1. Tracking all cgroup memory, it can quickly get expensive (tracking
> > > > to check for vm.soft_limit_ratio and for usage)
> > > 
> > > Not so expensive because we already tracks them all by default cgroup.
> > > Then, what we need is "fast" counter.
> > > Maybe percpu coutner (lib/percpu_counter.c) gives us enough codes for counting.
> > >
> > > Checking value ratio is ...how about "once per 1000 increment per cpu" or some ?
> > 
> > That is not true..we don't track them to default cgroup unless
> > memory.use_hiearchy is enabled in the root cgroup. 
> What I want to say is "the task which is not attached to user's cgroup is
> also under defaut cgroup, so we don't need additional hook"
> Not talking about hierarchy.
>

Since all the user pages are tracked in one or the other cgroup, the
total accounting is equal to total_lru_pages across all zones/nodes.
Your suggestion boils down to if total_lru_pages reaches a threshold,
do soft limit reclaim, instead of doing reclaim when there is
contention.. right?
 
> > To do what you
> > suggest, we have to iterate through all mem cgroups, which is not
> > desirable at all.
> 
> I don't say percpu counter should be within struct mem_cgroup.
> DEFINE_PERCPU(overall_usage);
> is enough, for example. (just an example.)
> 
> Or we already have vmstat (see vmstat.h:: global_page_state())
> 
> == si_meminfo()==
>  val->freeram = global_page_state(NR_FREE_PAGES);
> ==
> 
> It seems we have what we need already. (but usage from this includes usage comes
> from slab, page table, etc...)
> 
> 
> > > 
> > > > 2. Finding a good default for the sysctl (might not be so hard)
> > > > 
> > > I think some parameter like high-low watermark is good and we can find
> > > good value as
> > >   - low watermak .... max_memory - (sum of all zone->high) * 16 of memory.
> > >   - high watermark .... max_memory - (sum_of all zone->high) * 8
> > > (just an example but not so bad.)
> > >
> > 
> > OK..
> > 
> > [offtopic] I liked the per-mem cgroup watermark patches as well. I
> > think we should look at them later on, after soft limits and some other items.
> >  
> I'm glad if we can reuse logics added by softlimit by per memcg watermarks.
> 
> > > > Even today our influence on global LRU is very limited, only when we
> > > > come under reclaim, we do an additional step of seeing if we can get
> > > > memory from soft limit groups first.
> > > > 
> > > > (1) is a real concern.
> > > 
> > > Maybe yes. But all memcg will call "charge" "uncharge" codes so, problem is
> > > just "counter". I think percpu coutner works enough.
> > >
> > 
> > This scheme adds more overhead due to (1), we'll need a global counter
> > and need to protect it, which will serialize all res_counters. 
> > 
> It's not necessary. for example, reading vmstat doesn't need global lock.
>

It uses atomic values for accounting. 

-- 
	Balbir

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>