Re: [PATCH 3/4] sched: introduce synchronized idle injection

[Javi Merino]

On Tue, Nov 24, 2015 at 03:10:59AM -0800, Jacob Pan wrote:
> On Tue, 24 Nov 2015 11:09:56 +0100
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Tue, Nov 24, 2015 at 09:12:40AM +0000, Javi Merino wrote:
> > > On Mon, Nov 23, 2015 at 07:07:06PM +0100, Peter Zijlstra wrote:
> > > > On Mon, Nov 23, 2015 at 05:56:44PM +0000, Javi Merino wrote:
> > > > > I've tested this series on Juno (2xCortex-A57 4xCortex-A53).
> > > > > If you idle inject for 50% of the time, when I run 6 busy loops
> > > > > the scheduler sometimes keeps two of them in the same cpu while
> > > > > the another cpu is completely idle.  Without idle injection the
> > > > > scheduler does the sensible thing: put one busy loop in each
> > > > > CPU.  I'm running systemd and this only happens with
> > > > > CONFIG_SCHED_AUTOGROUP=y.  If I unset CONFIG_SCHED_AUTOGROUP,
> > > > > the tasks are spread across all cpus as usual.
> > > > 
> > > > That's not a plus for this patch though; but a bug report against
> > > > AUTOGROUP/cgroup muck, right?
> > > 
> > > The bug only happens when you apply this series and you set the
> > > system to idle inject 50% of the time.  SCHED_AUTOGROUP is already
> > > part of the kernel and behaves properly with the kernel as is.  I
> > > think that this patch should not introduce new bugs.
> > 
> > Ah, then I misunderstood your email, agreed.
> First of all, thanks for testing.
> Just trying to reproduce this. So let me understand your set up.
> - 8 cores in total?

- 6 cores in total

> - you first set 50% idle
> - then launch 6 busy loops

Correct

> How often you see this happen?

About 1 in 10.  Sometimes, even though it ends up doing the right
thing, it takes some time (>500ms) to move things around.  This never
happens with idle injection disabled (kernel.sched_cfs_idle_inject_pct
= 0).

Admittedly, the situation is a lot better than with 4.3.  When I first
tried this series on top of v4.3-rc6, the scheduler consistently
misbehaved with 50% idle injection, leaving up to 4 100% tasks on a
cpu while three cpus remain completely idle

Cheers,
Javi
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/