Re: [PATCH v1] mm:memcg: skip memcg of current in mem_cgroup_soft_limit_reclaim

On Fri 03-08-18 14:59:34, Zhaoyang Huang wrote:

On Fri, Aug 3, 2018 at 2:18 PM Michal Hocko [off-list ref] wrote:

quoted

On Fri 03-08-18 14:11:26, Zhaoyang Huang wrote:

quoted

On Fri, Aug 3, 2018 at 1:48 PM Zhaoyang Huang [off-list ref] wrote:

quoted

for the soft_limit reclaim has more directivity than global reclaim, we40960
have current memcg be skipped to avoid potential page thrashing.

The patch is tested in our android system with 2GB ram.  The case
mainly focus on the smooth slide of pictures on a gallery, which used
to stall on the direct reclaim for over several hundred
millionseconds. By further debugging, we find that the direct reclaim
spend most of time to reclaim pages on its own with softlimit set to
40960KB. I add a ftrace event to verify that the patch can help
escaping such scenario. Furthermore, we also measured the major fault
of this process(by dumpsys of android). The result is the patch can
help to reduce 20% of the major fault during the test.

I have asked already asked. Why do you use the soft limit in the first
place? It is known to cause excessive reclaim and long stalls.

It is required by Google for applying new version of android system.
There was such a mechanism called LMK in previous ANDROID version,
which will kill process when in memory contention like OOM does. I
think Google want to drop such rough way for reclaiming pages and turn
to memcg. They setup different memcg groups for different process of
the system and set their softlimit according to the oom_adj. Their
original purpose is to reclaim pages gentlely in direct reclaim and
kswapd. During the debugging process , it seems to me that memcg maybe
tunable somehow. At least , the patch works on our system.

Then the suggestion is to use v2 and the high limit. This is much less
disruptive method for pro-active reclaim. Really softlimit semantic is
established for many years and you cannot change it even when it sucks
for your workload. Others might depend on the traditional behavior.

I have tried to change the semantic in the past and there was a general
consensus that changing the semantic is just too risky. So it is nice
that it helps for your particular workload but this is not an upstream
material, I am sorry.

-- 
Michal Hocko
SUSE Labs