lkml.org 
[lkml]   [2009]   [Sep]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: tree rcu: call_rcu scalability problem?
On Wed, Sep 02, 2009 at 08:19:27AM -0700, Paul E. McKenney wrote:
> On Wed, Sep 02, 2009 at 02:27:56PM +0200, Nick Piggin wrote:
> > On Wed, Sep 02, 2009 at 11:48:35AM +0200, Nick Piggin wrote:
> > > Hi Paul,
> > >
> > > I'm testing out scalability of some vfs code paths, and I'm seeing
> > > a problem with call_rcu. This is a 2s8c opteron system, so nothing
> > > crazy.
> > >
> > > I'll show you the profile results for 1-8 threads:
> > >
> > > 1:
> > > 29768 total 0.0076
> > > 15550 default_idle 48.5938
> > > 1340 __d_lookup 3.6413
> > > 954 __link_path_walk 0.2559
> > > 816 system_call_after_swapgs 8.0792
> > > 680 kmem_cache_alloc 1.4167
> > > 669 dput 1.1946
> > > 591 __call_rcu 2.0521
> > >
> > > 2:
> > > 56733 total 0.0145
> > > 20074 default_idle 62.7313
> > > 3075 __call_rcu 10.6771
> > > 2650 __d_lookup 7.2011
> > > 2019 dput 3.6054
> > >
> > > 4:
> > > 98889 total 0.0253
> > > 21759 default_idle 67.9969
> > > 10994 __call_rcu 38.1736
> > > 5185 __d_lookup 14.0897
> > > 4475 dput 7.9911
>
> Four threads runs on one socket but 8 threads runs on two sockets,
> I take it?

Yes.


> > > 8:
> > > 170391 total 0.0437
> > > 31815 __call_rcu 110.4688
> > > 12958 dput 23.1393
> > > 10417 __d_lookup 28.3071
> > >
> > > Of course there are other scalability factors involved too, but
> > > __call_rcu is taking 54 times more CPU to do 8 times the amount
> > > of work from 1-8 threads, or a factor of 6.7 slowdown.
> > >
> > > This is with tree RCU.
> >
> > 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;
>
> Hmmm... That certainly is not the first list of code in call_rcu() that
> would come to mind...

It's weird. I *think* I read the asm right, but oprofile maybe is
not attributing the cost to the right instruction.


> > 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...
>
> The usual diagnosis would be false sharing.

Hmm that's possible yes.


> Hmmm... What is the workload? CPU-bound? If CONFIG_PREEMPT=n, I might
> expect interference from force_quiescent_state(), except that it should
> run only every few clock ticks. So this seems quite unlikely.

It's CPU bound and preempt=y.

Workload is just 8 processes running a loop of close(open("file$i")) as
I said though you probably won't be able to reproduce it on a vanilla
kernel.


> Could you please try padding the beginning and end of struct rcu_data
> with a few hundred bytes and rerunning? Just in case there is a shared
> per-CPU variable either before or after rcu_data in your memory layout?

OK I'll try that.



\
 
 \ /
  Last update: 2009-09-02 18:27    [from the cache]
©2003-2011 Jasper Spaans