Re: tree rcu: call_rcu scalability problem?

[Nick Piggin <npiggin@suse.de>]

On Wed, Sep 02, 2009 at 10:14:27PM -0700, Paul E. McKenney wrote:
> On Wed, Sep 02, 2009 at 09:17:44PM +0200, Peter Zijlstra wrote:
> > On Wed, 2009-09-02 at 14:27 +0200, Nick Piggin wrote:
> > 
> > > It seems like nearly 2/3 of the cost is here:
> > >         /* Add the callback to our list. */
> > >         *rdp->nxttail[RCU_NEXT_TAIL] = head; <<<
> > >         rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
> > > 
> > > In loading the pointer to the next tail pointer. If I'm reading the profile
> > > correctly. Can't see why that should be a probem though...
> > > 
> > > ffffffff8107dee0 <__call_rcu>: /* __call_rcu total: 320971 100.000 */
> > >    697  0.2172 :ffffffff8107dee0:       push   %r12
> > 
> > >    921  0.2869 :ffffffff8107df57:       push   %rdx
> > >    151  0.0470 :ffffffff8107df58:       popfq
> > > 183507 57.1725 :ffffffff8107df59:       mov    0x50(%rbx),%rax
> > >    995  0.3100 :ffffffff8107df5d:       mov    %rdi,(%rax)
> > 
> > I'd guess at popfq to be the expensive op here.. skid usually causes the
> > attribution to be a few ops down the line.
> 
> I believe that Nick's workload is routinely driving the number of
> callbacks queued on a given CPU above 10,000, which would provoke numerous
> (and possibly inlined) calls to force_quiescent_state().  Like about
> 400,000 such calls per second.  Hey, I was naively assuming that no one
> would see more than 10,000 callbacks queued on a single CPU unless there
> was some sort of major emergency underway, and coded accordingly.  ;-)
> 
> I offer the attached experimental (untested, might not even compile) patch.

Not only does it compile, but __call_rcu is now taking 1/10th the
cycles and absolute performance up nearly 20%. Looks like it is
now better than classic RCU.

I'll collect and post some more detailed numbers and profiles. Do
you want some new rcu trace results too?

Thanks,
Nick