On Wed 23-11-16 18:50:42, Balbir Singh wrote:

On 23/11/16 18:25, Michal Hocko wrote:

quoted

On Wed 23-11-16 15:36:51, Balbir Singh wrote:

quoted

In the absence of hotplug we use extra memory proportional to
(possible_nodes - online_nodes) * number_of_cgroups. PPC64 has a patch
to disable large consumption with large number of cgroups. This patch
adds hotplug support to memory cgroups and reverts the commit that
limited possible nodes to online nodes.

Balbir,
I have asked this in the previous version but there still seems to be a
lack of information of _why_ do we want this, _how_ much do we save on
the memory overhead on most systems and _why_ the additional complexity
is really worth it. Please make sure to add all this in the cover
letter.

The data is in the patch referred to in patch 3. The order of waste was
200MB for 400 cgroup directories enough for us to restrict possible_map
to online_map. These patches allow us to have a larger possible map and
allow onlining nodes not in the online_map, which is currently a restriction
on ppc64.

How common is to have possible_map >> online_map? If this is ppc64 then
what is the downside of keeping the current restriction instead?

A typical system that I use has about 100-150 directories, depending on the
number of users/docker instances/configuration/virtual machines. These numbers
will only grow as we pack more of these instances on them.

From a complexity view point, the patches are quite straight forward.

Well, I would like to hear more about that. {get,put}_online_memory
at random places doesn't sound all that straightforward to me.

-- 
Michal Hocko
SUSE Labs