Re: rcu: performance regression

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

On Tue, Jun 14, 2011 at 03:18:42PM +0200, Peter Zijlstra wrote:
> On Tue, 2011-06-14 at 15:14 +0200, Peter Zijlstra wrote:
> > On Tue, 2011-06-14 at 06:02 -0700, Paul E. McKenney wrote:
> > > > This commit also cause hackbench process mode performance dropping, and
> > > > Shaohua's patch do recovered this. But in hackbench testing, the vmstat
> > > > show context switch have some reduce. And perf tool show
> > > > root_domain->cpupri->prio_to_cpu[]->lock has contention with the commit.
> > > 
> > > Steven, Peter, would any of the recent fixes address this lock contention? 
> > 
> > No. People occasionally run into that thing, but given the constraints
> > on SCHED_FIFO SMP (run the n highest priority tasks on m cpus; n<=m)
> > there isn't really much we can do.
> > 
> > IIRC Chris Mason once proposed a lockless version, but that opened up
> > race windows, Steve might know the details better.
> 
> Note that even a lockless version will die of cacheline contention.

Hmmm...  This is going to be fun for the -rt guys given the mapping of
softirq and irq to kthreads, but then again, perhaps people will
confine use of -rt to smaller systems.

Thankfully, Shaohua's patch is looking good so far, give or take a
couple of future RCU items, but then again, those are in the future
rather than here and now.

							Thanx, Paul