memcg: gix memory accounting scalability in shrink_page_list
authorTim Chen <tim.c.chen@linux.intel.com>
Tue, 31 Jul 2012 23:46:08 +0000 (16:46 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 1 Aug 2012 01:42:49 +0000 (18:42 -0700)
I noticed in a multi-process parallel files reading benchmark I ran on a 8
socket machine, throughput slowed down by a factor of 8 when I ran the
benchmark within a cgroup container.  I traced the problem to the
following code path (see below) when we are trying to reclaim memory from
file cache.  The res_counter_uncharge function is called on every page
that's reclaimed and created heavy lock contention.  The patch below
allows the reclaimed pages to be uncharged from the resource counter in
batch and recovered the regression.

Tim

     40.67%           usemem  [kernel.kallsyms]                   [k] _raw_spin_lock
                      |
                      --- _raw_spin_lock
                         |
                         |--92.61%-- res_counter_uncharge
                         |          |
                         |          |--100.00%-- __mem_cgroup_uncharge_common
                         |          |          |
                         |          |          |--100.00%-- mem_cgroup_uncharge_cache_page
                         |          |          |          __remove_mapping
                         |          |          |          shrink_page_list
                         |          |          |          shrink_inactive_list
                         |          |          |          shrink_mem_cgroup_zone
                         |          |          |          shrink_zone
                         |          |          |          do_try_to_free_pages
                         |          |          |          try_to_free_pages
                         |          |          |          __alloc_pages_nodemask
                         |          |          |          alloc_pages_current

Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c

index e37e6872509064cd332b86dff1416a5e48d7580f..8d01243d9560e0ea8d8a04cf51cd4087ed8b280d 100644 (file)
@@ -687,6 +687,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 
        cond_resched();
 
+       mem_cgroup_uncharge_start();
        while (!list_empty(page_list)) {
                enum page_references references;
                struct address_space *mapping;
@@ -953,6 +954,7 @@ keep:
 
        list_splice(&ret_pages, page_list);
        count_vm_events(PGACTIVATE, pgactivate);
+       mem_cgroup_uncharge_end();
        *ret_nr_dirty += nr_dirty;
        *ret_nr_writeback += nr_writeback;
        return nr_reclaimed;