[lkml]   [2018]   [Mar]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCH 00/10] Improve shrink_slab() scalability (old complexity was O(n^2), new is O(n))
Imagine a big node with many cpus, memory cgroups and containers.
Let we have 200 containers, and every container has 10 mounts
and 10 cgroups. All container tasks don't touch foreign containers

In case of global reclaim, a task has to iterate all over the memcgs
and to call all the memcg-aware shrinkers for all of them. This means,
the task has to visit 200 * 10 = 2000 shrinkers for every memcg,
and since there are 2000 memcgs, the total calls of do_shrink_slab()
are 2000 * 2000 = 4000000.

4 million calls are not a number operations, which can takes 1 cpu cycle.
E.g., super_cache_count() accesses at least two lists, and makes arifmetical
calculations. Even, if there are no charged objects, we do these calculations,
and replaces cpu caches by read memory. I observed nodes spending almost 100%
time in kernel, in case of intensive writing and global reclaim. Even if
there is no writing, the iterations just waste the time, and slows reclaim down.

Let's see the small test below:
$echo 1 > /sys/fs/cgroup/memory/memory.use_hierarchy
$mkdir /sys/fs/cgroup/memory/ct
$echo 4000M > /sys/fs/cgroup/memory/ct/memory.kmem.limit_in_bytes
$for i in `seq 0 4000`; do mkdir /sys/fs/cgroup/memory/ct/$i; echo $$ > /sys/fs/cgroup/memory/ct/$i/cgroup.procs; mkdir -p s/$i; mount -t tmpfs $i s/$i; touch s/$i/file; done

Then, let's see drop caches time (4 sequential calls):
$time echo 3 > /proc/sys/vm/drop_caches
0.00user 6.80system 0:06.82elapsed 99%CPU
0.00user 4.61system 0:04.62elapsed 99%CPU
0.00user 4.61system 0:04.61elapsed 99%CPU
0.00user 4.61system 0:04.61elapsed 99%CPU

Last three calls don't actually shrink something. So, the iterations
over slab shrinkers take 4.61 seconds. Not so good for scalability.

The patchset solves the problem with following actions:
1)Assign id to every registered memcg-aware shrinker.
2)Maintain per-memcgroup bitmap of memcg-aware shrinkers,
and set a shrinker-related bit after the first element
is added to lru list (also, when removed child memcg
elements are reparanted).
3)Split memcg-aware shrinkers and !memcg-aware shrinkers,
and call a shrinker if its bit is set in memcg's shrinker
(Also, there is a functionality to clear the bit, after
last element is shrinked).

This gives signify performance increase. The result after patchset is applied:

$time echo 3 > /proc/sys/vm/drop_caches
0.00user 0.93system 0:00.94elapsed 99%CPU
0.00user 0.00system 0:00.01elapsed 80%CPU
0.00user 0.00system 0:00.01elapsed 80%CPU
0.00user 0.00system 0:00.01elapsed 81%CPU
(4.61s/0.01s = 461 times faster)

Currenly, all memcg-aware shrinkers are implemented via list_lru.
The only exception is XFS cached objects backlog (which is completelly
no memcg-aware, but pretends to be memcg-aware). See
xfs_fs_nr_cached_objects() and xfs_fs_free_cached_objects() for
the details. It seems, this can be reworked to fix this lack.

So, the patchset makes shrink_slab() of less complexity and improves
the performance in such types of load I pointed. This will give a profit
in case of !global reclaim case, since there also will be less do_shrink_slab()

This patchset is made against linux-next.git tree.


Kirill Tkhai (10):
mm: Assign id to every memcg-aware shrinker
mm: Maintain memcg-aware shrinkers in mcg_shrinkers array
mm: Assign memcg-aware shrinkers bitmap to memcg
fs: Propagate shrinker::id to list_lru
list_lru: Add memcg argument to list_lru_from_kmem()
list_lru: Pass dst_memcg argument to memcg_drain_list_lru_node()
list_lru: Pass lru argument to memcg_drain_list_lru_node()
mm: Set bit in memcg shrinker bitmap on first list_lru item apearance
mm: Iterate only over charged shrinkers during memcg shrink_slab()
mm: Clear shrinker bit if there are no objects related to memcg

fs/super.c | 8 +
include/linux/list_lru.h | 3
include/linux/memcontrol.h | 20 +++
include/linux/shrinker.h | 9 +
mm/list_lru.c | 65 ++++++--
mm/memcontrol.c | 7 +
mm/vmscan.c | 337 ++++++++++++++++++++++++++++++++++++++++++--
mm/workingset.c | 6 +
8 files changed, 418 insertions(+), 37 deletions(-)

Signed-off-by: Kirill Tkhai <>

 \ /
  Last update: 2018-03-21 14:25    [W:0.308 / U:0.428 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site