Date: Thu, 28 Apr 2011 08:57:51 +0900
From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
To: Ying Han <yinghan@google.com>
Cc: "linux-mm@kvack.org" <linux-mm@kvack.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "akpm@linux-foundation.org" <akpm@linux-foundation.org>,
        "nishimura@mxp.nes.nec.co.jp" <nishimura@mxp.nes.nec.co.jp>,
        "balbir@linux.vnet.ibm.com" <balbir@linux.vnet.ibm.com>
Subject: Re: [PATCHv2] memcg: reclaim memory from node in round-robin
Message-Id: <20110428085751.fd478fe8.kamezawa.hiroyu@jp.fujitsu.com>
In-Reply-To: <BANLkTinx+4zXaO3rhHRUzr3m-K-2_NMTQw@mail.gmail.com>
References: <20110427165120.a60c6609.kamezawa.hiroyu@jp.fujitsu.com>
	<BANLkTinx+4zXaO3rhHRUzr3m-K-2_NMTQw@mail.gmail.com>
Organization: FUJITSU Co. LTD.
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 7836
Lines: 197

On Wed, 27 Apr 2011 10:33:43 -0700
Ying Han <yinghan@google.com> wrote:

> On Wed, Apr 27, 2011 at 12:51 AM, KAMEZAWA Hiroyuki
> <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > I changed the logic a little and add a filter for skipping nodes.
> > With large NUMA, tasks may under cpuset or mempolicy and the usage of memory
> > can be unbalanced. So, I think a filter is required.
> 
> Thank you.
> 
> >
> > ==
> > Now, memory cgroup's direct reclaim frees memory from the current node.
> > But this has some troubles. In usual, when a set of threads works in
> > cooperative way, they are tend to on the same node. So, if they hit
> > limits under memcg, it will reclaim memory from themselves, it may be
> > active working set.
> >
> > For example, assume 2 node system which has Node 0 and Node 1
> > and a memcg which has 1G limit. After some work, file cacne remains and
> > and usages are
> >   Node 0:  1M
> >   Node 1:  998M.
> >
> > and run an application on Node 0, it will eats its foot before freeing
> > unnecessary file caches.
> >
> > This patch adds round-robin for NUMA and adds equal pressure to each
> > node. When using cpuset's spread memory feature, this will work very well.
> >
> >
> > From: Ying Han <yinghan@google.com>
> > Signed-off-by: Ying Han <yinghan@google.com>
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
> >
> > Changelog v1->v2:
> >  - fixed comments.
> >  - added a logic to avoid scanning unused node.
> >
> > ---
> >  include/linux/memcontrol.h |    1
> >  mm/memcontrol.c            |   98 ++++++++++++++++++++++++++++++++++++++++++---
> >  mm/vmscan.c                |    9 +++-
> >  3 files changed, 101 insertions(+), 7 deletions(-)
> >
> > Index: memcg/include/linux/memcontrol.h
> > ===================================================================
> > --- memcg.orig/include/linux/memcontrol.h
> > +++ memcg/include/linux/memcontrol.h
> > @@ -108,6 +108,7 @@ extern void mem_cgroup_end_migration(str
> >  */
> >  int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg);
> >  int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg);
> > +int mem_cgroup_select_victim_node(struct mem_cgroup *memcg);
> >  unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
> >                                       struct zone *zone,
> >                                       enum lru_list lru);
> > Index: memcg/mm/memcontrol.c
> > ===================================================================
> > --- memcg.orig/mm/memcontrol.c
> > +++ memcg/mm/memcontrol.c
> > @@ -237,6 +237,11 @@ struct mem_cgroup {
> >         * reclaimed from.
> >         */
> >        int last_scanned_child;
> > +       int last_scanned_node;
> > +#if MAX_NUMNODES > 1
> > +       nodemask_t      scan_nodes;
> > +       unsigned long   next_scan_node_update;
> > +#endif
> >        /*
> >         * Should the accounting and control be hierarchical, per subtree?
> >         */
> > @@ -650,18 +655,27 @@ static void mem_cgroup_soft_scan(struct
> >        this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_SOFT_SCAN], val);
> >  }
> >
> > +static unsigned long
> > +mem_cgroup_get_zonestat_node(struct mem_cgroup *mem, int nid, enum lru_list idx)
> > +{
> > +       struct mem_cgroup_per_zone *mz;
> > +       u64 total;
> > +       int zid;
> > +
> > +       for (zid = 0; zid < MAX_NR_ZONES; zid++) {
> > +               mz = mem_cgroup_zoneinfo(mem, nid, zid);
> > +               total += MEM_CGROUP_ZSTAT(mz, idx);
> > +       }
> > +       return total;
> > +}
> >  static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
> >                                        enum lru_list idx)
> >  {
> > -       int nid, zid;
> > -       struct mem_cgroup_per_zone *mz;
> > +       int nid;
> >        u64 total = 0;
> >
> >        for_each_online_node(nid)
> > -               for (zid = 0; zid < MAX_NR_ZONES; zid++) {
> > -                       mz = mem_cgroup_zoneinfo(mem, nid, zid);
> > -                       total += MEM_CGROUP_ZSTAT(mz, idx);
> > -               }
> > +               total += mem_cgroup_get_zonestat_node(mem, nid, idx);
> >        return total;
> >  }
> >
> > @@ -1471,6 +1485,77 @@ mem_cgroup_select_victim(struct mem_cgro
> >        return ret;
> >  }
> >
> > +#if MAX_NUMNODES > 1
> > +
> > +/*
> > + * Update nodemask always is not very good. Even if we have empty
> > + * list, or wrong list here, we can start from some node and traverse all nodes
> > + * based on zonelist. So, update the list loosely once in 10 secs.
> > + *
> > + */
> > +static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
> > +{
> > +       int nid;
> > +
> > +       if (time_after(mem->next_scan_node_update, jiffies))
> > +               return;
> > +
> > +       mem->next_scan_node_update = jiffies + 10*HZ;
> > +       /* make a nodemask where this memcg uses memory from */
> > +       mem->scan_nodes = node_states[N_HIGH_MEMORY];
> > +
> > +       for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
> > +
> > +               if (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_FILE) ||
> > +                   mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_FILE))
> > +                       continue;
> > +
> > +               if (total_swap_pages &&
> > +                   (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_ANON) ||
> > +                    mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_ANON)))
> > +                       continue;
> > +               node_clear(nid, mem->scan_nodes);
> > +       }
> > +
> > +}
> > +
> > +/*
> > + * Selecting a node where we start reclaim from. Because what we need is just
> > + * reducing usage counter, start from anywhere is O,K. Considering
> > + * memory reclaim from current node, there are pros. and cons.
> > + *
> > + * Freeing memory from current node means freeing memory from a node which
> > + * we'll use or we've used. So, it may make LRU bad. And if several threads
> > + * hit limits, it will see a contention on a node. But freeing from remote
> > + * node means more costs for memory reclaim because of memory latency.
> > + *
> > + * Now, we use round-robin. Better algorithm is welcomed.
> > + */
> > +int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
> > +{
> > +       int node;
> > +
> > +       mem_cgroup_may_update_nodemask(mem);
> > +       node = mem->last_scanned_node;
> > +
> > +       node = next_node(node, mem->scan_nodes);
> > +       if (node == MAX_NUMNODES) {
> > +               node = first_node(mem->scan_nodes);
> > +               if (unlikely(node == MAX_NUMNODES))
> > +                       node = numa_node_id();
> not sure about this logic, is that possible we reclaim from a node
> with all "unreclaimable" pages (based on the
> mem_cgroup_may_update_nodemask check).
> If i missed anything here, it would be helpful to add comment.
> 

What I'm afraid here is when a user uses very small memcg,
all pages on the LRU may be isolated or all usages are in per-cpu cache
of memcg or because of task-migration between memcg, it hits limit before
having any pages on LRU.....I think there is possible corner cases which
can cause hang.

ok, will add comment.

Thanks,
-Kame


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/