Re: [PATCH RFC tip/core/rcu] Parallelize and economize NOCB kthread wakeups

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Jul 04, 2014 at 08:01:34AM +0200, Mike Galbraith wrote:
> On Thu, 2014-07-03 at 22:05 -0700, Paul E. McKenney wrote: 
> > On Fri, Jul 04, 2014 at 05:23:56AM +0200, Mike Galbraith wrote:
> 
> > > Turn it on and don't worry about it is exactly what distros want the
> > > obscure feature with very few users to be.  Last time I did a drive-by,
> > > my boxen said I should continue to worry about it ;-)
> > 
> > Yep, which is the reason for the patch on the last email.
> > 
> > Then again, exactly which feature and which reason for worry?
> 
> NO_HZ_FULL.  I tried ALL a while back, box instantly called me an idiot.
> Maybe that has improved since, dunno.

Ah, I was thinking in terms of RCU_CPU_NOCB.

> Last drive-by I didn't do much overhead measurement, stuck mostly
> functionality, and it still had rough edges that enterprise users may
> not fully appreciate.  Trying to let 60 of 64 cores do 100% compute
> showed some cores having a hard time entering tickless at all, and
> ~200us spikes that I think are due to tick losing skew.. told Frederic
> I'd take a peek at that, but haven't had time yet.  There were other
> known things as well, like timers and workqueues for which there are
> patches floating around.  All in all, it was waving the men at work
> sign, pointing at the "Say N" by the config option, and suggesting that
> ignoring that would not be the cleverest of moves.

Well, I am not going to join a debate on Kconfig default selection.  ;-)
I will say that two years ago, setting NO_HZ_FULL=y by default would
have been insane.  Perhaps soon it will be a no-brainer, and I am of
course trying to bring that day closer.  Right now it is of course a
judgment call.

							Thanx, Paul