Re: [RFC PATCH 1/9] sched,cgroup: Add interface for latency-nice

[Joel Fernandes]

Hi Qais!

On Mon, Apr 20, 2020 at 12:47:29PM +0100, Qais Yousef wrote:
> On 04/18/20 12:01, Joel Fernandes wrote:
> > > What's missing is the per-taskgroup implementation, at least from the
> > > standpoint of ACK.
> > >
> > > The (mainline) EAS use-case for latency nice is already in ACK
> > > (android-5.4):
> > >
> > > https://android.googlesource.com/kernel/common/+/760b82c9b88d2c8125abfc5f732cc3cd460b2a54
> > 
> > Yes, I was aware of this. But if we use task groups, then the
> > transition from schedtune -> uclamp means now the tasks that use
> > uclamp would also be subjected to cpu.shares. That's why we were
> > looking into the per-task interface and glad there's some work on this
> > already done.
> 
> Hmm uclamp doesn't do anything with cpu.shares. I assume this is some
> implementation detail at your end? IOW, you don't have to use cpu.shares to use
> uclamp.

Right, it is a ChromeOS-specific issue. We have CONFIG_FAIR_GROUP_SCHED
enabled in the kernel for container workloads. However there are CGroups of
tasks that used "schedtune" CGroup interface before to provide util clamping
like behavior. We are now migrating these to the upstream util-clamp.

We can't disable CONFIG_FAIR_GROUP_SCHED because that would break the
container workloads.

So we have to use the per-process interface of util clamp.

If we used the CGroups interface of util clamping, we would get the
cpu.shares as well since the CGroup interface comes with shares. There's no
way to avoid being subject to cpu.shares (that I'm aware off anyway).

> Although there should be few tasks in the system that need the latency-nice, so
> I prefer the per-task interface rather than lump everything in a cgroup. Though
> there could be valid use cases for the latter.

Yes, with either interface, we need something like latency_nice to indicate
that the task is low-latency (something we used for a number of years with
the out-of-tree schedtune).

thanks!

 - Joel


> 
> Thanks
> 
> --
> Qais Yousef