Re: [PATCH v2] sched: let __sched_period() use rq's nr_running

[Mike Galbraith]

On Mon, 2015-07-13 at 17:29 +0900, Byungchul Park wrote:
> On Mon, Jul 13, 2015 at 09:07:01AM +0200, Mike Galbraith wrote:
> > On Mon, 2015-07-13 at 09:56 +0900, Byungchul Park wrote:
> > 
> > > and i agree with that it makes latency increase for non-grouped tasks.
> > 
> > It's not only a latency hit for the root group, it's across the board.
> > 
> > I suspect an overloaded group foo/bar/baz would prefer small slices over
> > a large wait as well. I certainly wouldn't want my root group taking the
> > potentially huge hits that come with stretching period to accommodate an
> > arbitrarily overloaded /foo/bar/baz.
> 
> hello, mike :)
> 
> ok, then, do you think that the period have to be stretched by the number of
> rq's sched entity(e.i. rq->cfs.nr_running)? if it is done with rq->cfs.nr_running,
> as you can guess, leaf sched entities(e.i. tasks) can have much smaller slice
> than sysctl_sched_min_granularity. and some code using sysctl_sched_min_granularity
> need to be fixed in addition.

The only choice is to give a small slice frequently or a large slice
infrequently.  Increasing spread for the entire world to accommodate a
massive overload of a small share group in some hierarchy is just not a
viable option.

> anyway, current code looks broken since it stretching with local cfs's nr_running.
> IMHO, it should be stretched with rq->*cfs.nr_running* though leaf tasks can have
> very small slice, or it should be stretched with rq->*nr_running* to ensure that
> any task can have a slice which can be comparable to sysctl_sched_min_granularity.
> 
> what do you think about this concern?

It seems to work fine.  Just say "oh hell no" to hierarchies, and if you
think slices are too small, widen latency_ns a bit to get what you want
to see on your box.  Computing latency target bottom up and cumulatively
is a very bad idea, that lets one nutty group dictate latency for all.  

Something else to keep in mind when fiddling is that FAIR_SLEEPERS by
definition widens spread, effectively doubling our latency target, as
the thing it is defined by is that latency target.  We need short term
fairness so sleepers can perform when facing a world full of hogs, but
the last thing we need is short term being redefined to a week or so ;-)

	-Mike

--
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/