Re: [PATCH] net: use synchronize_rcu_expedited()

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

On Tue, May 24, 2011 at 05:52:44PM +0200, Eric Dumazet wrote:
> Le mardi 24 mai 2011 à 08:44 -0700, Paul E. McKenney a écrit :
> > On Tue, May 24, 2011 at 11:07:32AM +0200, Eric Dumazet wrote:
> > > synchronize_rcu() is very slow in various situations (HZ=100,
> > > CONFIG_NO_HZ=y, CONFIG_PREEMPT=n)
> > > 
> > > Extract from my (mostly idle) 8 core machine :
> > > 
> > >  synchronize_rcu() in 99985 us
> > >  synchronize_rcu() in 79982 us
> > >  synchronize_rcu() in 87612 us
> > >  synchronize_rcu() in 79827 us
> > >  synchronize_rcu() in 109860 us
> > >  synchronize_rcu() in 98039 us
> > >  synchronize_rcu() in 89841 us
> > >  synchronize_rcu() in 79842 us
> > >  synchronize_rcu() in 80151 us
> > >  synchronize_rcu() in 119833 us
> > >  synchronize_rcu() in 99858 us
> > >  synchronize_rcu() in 73999 us
> > >  synchronize_rcu() in 79855 us
> > >  synchronize_rcu() in 79853 us
> > > 
> > > 
> > > When we hold RTNL mutex, we would like to spend some cpu cycles but not
> > > block too long other processes waiting for this mutex.
> > > 
> > > We also want to setup/dismantle network features as fast as possible at
> > > boot/shutdown time.
> > > 
> > > This patch makes synchronize_net() call the expedited version if RTNL is
> > > locked.
> > > 
> > > synchronize_rcu_expedited() typical delay is about 20 us on my machine.
> > > 
> > >  synchronize_rcu_expedited() in 18 us
> > >  synchronize_rcu_expedited() in 18 us
> > >  synchronize_rcu_expedited() in 18 us
> > >  synchronize_rcu_expedited() in 18 us
> > >  synchronize_rcu_expedited() in 20 us
> > >  synchronize_rcu_expedited() in 16 us
> > >  synchronize_rcu_expedited() in 20 us
> > >  synchronize_rcu_expedited() in 18 us
> > >  synchronize_rcu_expedited() in 18 us
> > 
> > Cool!!!
> > 
> > Just out of curiosity, how many CPUs does your system have?
> 
> 16 (2x4x2)  [ processor.max_cstate=1 ]
> 
> I am now trying to optimize rcu_barrier(), if you have an idea to get an
> expedited version as well ?
> 
> We can see in following trace 3 groups, spaced by one jiffie (HZ=100)
> 
> Maybe we can avoid sending a call_rcu() to a cpu that has no pending rcu
> work ?

Might make sense, though most of the gains would need to come from
kicking the grace-period machinery hard in order to make it go faster.

Interesting -- I will give this some thought.

							Thanx, Paul

> [  835.189996] cpu0 synchronize_rcu_expedited() in 30 us 
>    -> begin rcu_barrier() immediately
> [  835.259702] cpu15 rcu_barrier_callback()
> [  835.259705] cpu14 rcu_barrier_callback()
> [  835.259708] cpu7 rcu_barrier_callback()
> [  835.259711] cpu12 rcu_barrier_callback()
> [  835.259714] cpu8 rcu_barrier_callback()
> [  835.259716] cpu1 rcu_barrier_callback()
> [  835.259719] cpu0 rcu_barrier_callback()
> 
> [  835.269691] cpu13 rcu_barrier_callback()
> [  835.269695] cpu11 rcu_barrier_callback()
> [  835.269698] cpu5 rcu_barrier_callback()
> [  835.269700] cpu6 rcu_barrier_callback()
> [  835.269702] cpu10 rcu_barrier_callback()
> [  835.269705] cpu3 rcu_barrier_callback()
> [  835.269707] cpu2 rcu_barrier_callback()
> 
> [  835.279687] cpu4 rcu_barrier_callback()
> [  835.279689] cpu9 rcu_barrier_callback()
> [  835.279744] cpu0 rcu_barrier() in 89499 us
> 
> Thanks
> 
>