Re: [PATCH] mm/memcg: support control THP behaviour in cgroup

[CGEL <cgel.zte@gmail.com>]

On Tue, May 10, 2022 at 03:36:34PM +0200, Michal Hocko wrote:
> On Tue 10-05-22 11:52:51, CGEL wrote:
> > On Tue, May 10, 2022 at 12:00:04PM +0200, Michal Hocko wrote:
> > > On Tue 10-05-22 01:43:38, CGEL wrote:
> > > > On Mon, May 09, 2022 at 01:48:39PM +0200, Michal Hocko wrote:
> > > > > On Mon 09-05-22 11:26:43, CGEL wrote:
> > > > > > On Mon, May 09, 2022 at 12:00:28PM +0200, Michal Hocko wrote:
> > > > > > > On Sat 07-05-22 02:05:25, CGEL wrote:
> > > > > > > [...]
> > > > > > > > If there are many containers to run on one host, and some of them have high
> > > > > > > > performance requirements, administrator could turn on thp for them:
> > > > > > > > # docker run -it --thp-enabled=always
> > > > > > > > Then all the processes in those containers will always use thp.
> > > > > > > > While other containers turn off thp by:
> > > > > > > > # docker run -it --thp-enabled=never
> > > > > > > 
> > > > > > > I do not know. The THP config space is already too confusing and complex
> > > > > > > and this just adds on top. E.g. is the behavior of the knob
> > > > > > > hierarchical? What is the policy if parent memcg says madivise while
> > > > > > > child says always? How does the per-application configuration aligns
> > > > > > > with all that (e.g. memcg policy madivise but application says never via
> > > > > > > prctl while still uses some madvised - e.g. via library).
> > > > > > >
> > > > > > 
> > > > > > The cgroup THP behavior is align to host and totally independent just likes
> > > > > > /sys/fs/cgroup/memory.swappiness. That means if one cgroup config 'always'
> > > > > > for thp, it has no matter with host or other cgroup. This make it simple for
> > > > > > user to understand or control.
> > > > > 
> > > > > All controls in cgroup v2 should be hierarchical. This is really
> > > > > required for a proper delegation semantic.
> > > > >
> > > > 
> > > > Could we align to the semantic of /sys/fs/cgroup/memory.swappiness?
> > > > Some distributions like Ubuntu is still using cgroup v1.
> > > 
> > > cgroup v1 interface is mostly frozen. All new features are added to the
> > > v2 interface.
> > >
> > 
> > So what about we add this interface to cgroup v2?
> 
> Can you come up with a sane hierarchical behavior?
> 
> [...]
> > > > For micro-service architecture, the application in one container is not a
> > > > set of loosely tight processes, it's aim at provide one certain service,
> > > > so different containers means different service, and different service
> > > > has different QoS demand. 
> > > 
> > > OK, if they are tightly coupled you could apply the same THP policy by
> > > an existing prctl interface. Why is that not feasible. As you are noting
> > > below...
> > > 
> > > >     5.containers usually managed by compose software, which treats container as
> > > > base management unit;
> > > 
> > > ..so the compose software can easily start up the workload by using prctl
> > > to disable THP for whatever workloads it is not suitable for.
> > 
> > prctl(PR_SET_THP_DISABLE..) can not be elegance to support the semantic we
> > need. If only some containers needs THP, other containers and host do not need
> > THP. We must set host THP to always first, and call prctl() to close THP for
> > host tasks and other containers one by one,
> 
> It might not be the most elegant solution but it should work.
> Maintaining user interfaces for ever has some cost and the THP
> configuration space is quite large already. So I would rather not add
> more complication in unless that is absolutely necessary.
>

By the way, should we let prctl() support PR_SET_THP_ALWAYS? Just likes
PR_TASK_PERF_EVENTS_DISABLE and PR_TASK_PERF_EVENTS_ENABLE. This would
make it simpler to let certain process use THP while others not use.

> > in this process some tasks that start before we call prctl() may
> > already use THP with no need.
> 
> As long as all those processes have a common ancestor I do not see how
> that would be possible.
> 
> -- 
> Michal Hocko
> SUSE Labs