Re: [PATCH 4/4] nohz: Set isolcpus when nohz_full is set

[Mike Galbraith <umgwanakikbuti@gmail.com>]

On Sat, 2015-05-16 at 15:39 -0400, Sasha Levin wrote:
> On 05/06/2015 12:04 PM, Frederic Weisbecker wrote:
> > From: Chris Metcalf <cmetcalf@ezchip.com>
> > 
> > nohz_full is only useful with isolcpus also set, since otherwise the
> > scheduler has to run periodically to try to determine whether to steal
> > work from other cores.
> > 
> > Accordingly, when booting with nohz_full=xxx on the command line, we
> > should act as if isolcpus=xxx was also set, and set (or extend) the
> > isolcpus set to include the nohz_full cpus.
> > 
> > Acked-by: Mike Galbraith <umgwanakikbuti@gmail.com> ["thumbs up!"]
> > Acked-by: Rik van Riel <riel@redhat.com>
> > Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> > Signed-off-by: Chris Metcalf <cmetcalf@ezchip.com>
> > Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
> > Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
> > Cc: Mike Galbraith <umgwanakikbuti@gmail.com>
> > Cc: Ingo Molnar <mingo@redhat.com>
> > Cc: Rik van Riel <riel@redhat.com>
> > Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
> 
> Hi folks,
> 
> I've noticed a regression in my testing a few days ago and bisected it down to
> this patch. I was seeing frequent soft lockups/RCU lockups and the load of the
> testing VMs would go beyond 400-500 (on 32 VCPU guests) - note I'm booting them
> with nohz_full=1-27.
> 
> This patch sort of explains the behaviour I was seeing now: most of the cores
> are no longer being used by the scheduler, and the remaining cores can't deal
> with the load imposed on them which results in "lockups" which are really just
> the CPUs being unable to keep up.
> 
> I always thought that nohz_full without isolcpus meant that the cores would
> be available to the scheduler, but it won't interfere if there is one task
> running on them. It seems that this patch changed that behaviour.
> 
> Did I misunderstand that?

Yeah, tying nohz_full set to isolcpus set up an initial condition that
you have to tear down with cpusets if you want those cpus returned to
the general purpose pool.  I had considered the kernel setting initial
state to be an optimization, but have reconsidered.

You may have misunderstood somewhat though, if you do not explicitly
isolate the nohz_full set from the scheduler via either isolcpus or
cpusets, there is no exclusion from load balancing, the scheduler may
place additional tasks on a nohz_full cpu to resolve load imbalance.

Given that kernel initiated association to isolcpus, a user turning
NO_HZ_FULL_ALL on had better not have much generic load to manage.  If
he/she does not have CPUSETS enabled, or should Rik's patch rendering
isolcpus immutable be merged, he/she would quickly discover that the
generic box has been transformed into an utterly inflexible specialist
incapable of performing mundane tasks ever again.

	-Mike