[Ksummit-2013-discuss] [ATTEND] memcg topics, cgroups conversion to single agent, other MM stuff

[Michal Hocko]

Hi,
I am one of the memory cgroups maintainer and am active also in general
memory management area which is a expertize I can bring to the KS which
I would like to attend this year.

There are many topics to discuss on memcg front - e.g. a new
lowlimit/reclaim_threshold knobs which would allow for guarantee of
minimum memory kept for a group/hierarchy. We originally planned to
provide similar functionality via an existing soft_limit knob but this
has been shot down by Tejun because the interface wouldn't be clean
enough.
I would also like to touch memcg dirty pages accounting and how to
proceed with dirty memory throttling. First part is on a good way to get
merged in a foreseeable future but the later one is rather complicated
(do we want to have per group/hierarchy flusher threads?)
Other topic would be userspace global OOM handling. Google would like
to reuse oom_control interface which we have for !root memcgs. The idea
sounds natural but we should consider other options (allow to plug
modules to handle OOM conditions or even allow a scriptable interface).

I am also taking care of cgroups inside SUSE so I would be really
interested in discussing future plans of cgroup unified hierarchy. While
the idea sounds good in general, convergence to a single management tool
which would hide cgroupfs sounds really intrusive and it has some
potential to break our users' use cases. "It's all ours" claim by
systemd just amplifies our worries.
-- 
Michal Hocko
SUSE Labs