Re: [PATCH] sched: Make schedstats a runtime tunable that is disabled by default

From: Ingo Molnar
Date: Mon Jan 25 2016 - 10:46:59 EST



* Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Mon, Jan 25, 2016 at 01:39:44PM +0000, Mel Gorman wrote:
> > On Mon, Jan 25, 2016 at 12:26:06PM +0100, Peter Zijlstra wrote:
> > > On Mon, Jan 25, 2016 at 10:05:31AM +0000, Mel Gorman wrote:
> > > > schedstats is very useful during debugging and performance tuning but it
> > > > incurs overhead. As such, even though it can be disabled at build time,
> > > > it is often enabled as the information is useful. This patch adds a
> > > > kernel command-line and sysctl tunable to enable or disable schedstats on
> > > > demand. It is disabled by default as someone who knows they need it can
> > > > also learn to enable it when necessary.
> > >
> > > So the reason its often enabled in distro configs is (IIRC) that it
> > > enables trace_sched_stat_{wait,sleep,iowait,blocked}().
> > >
> > > I've not looked at the details of this patch, but I suspect this patch
> > > would make these tracepoints available but non-functional unless you
> > > poke the magic button.
> > >
> >
> > It's potentially slightly worse than that. The tracepoints are available,
> > functional but produce garbage unless the magic button is poked and do
> > a lot of useful work producing that garbage. I missed a few hunks that
> > are included below. With this, the tracepoints will exist but unless the
> > magic button is poked, they'll never fire. Considering the paths
> > affected, this will require retesting but if it's ok, would you be ok in
> > general with a patch like this that forces a button to be pushed if
> > the user is doing performance analysis?
>
> Its rather unintuitive and error prone semantics :/
>
> Ideally we'd auto-magically enable the magic knob if any of these
> affected tracepoints become active. Or alternatively fail to enable the
> tracepoints (which would then get us people going: 'WTF this used to
> work').
>
> One of the things on my TODO is look at how much of sched_stat is
> required for these tracepoints and see if we can enable just that
> (hopefully) little bit, while not doing the rest of the accounting.
>
> Of course, it'll be our luck that tracking the data for these
> tracepoints is the most expensive part of schedstats ...
>
> Ingo?

IIRC it needed only a small subset of schedstats to make those tracepoints work.

We already have too much overhead in the scheduler as-is - and the extra cache
footprint does not even show on the typically cache-rich enterprise CPUs most of
the scalability testing goes on.

My minimum requirement for such runtime enablement would be to make it entirely
static-branch patched and triggered at the call sites as well - not hidden inside
schedstat functions.

Thanks,

Ingo