Re: [PATCH] rwsem: reduce spinlock contention in wakeup code path

[Ingo Molnar <mingo@kernel.org>]

* Tim Chen <tim.c.chen@linux.intel.com> wrote:

> On Fri, 2013-09-27 at 12:39 -0700, Davidlohr Bueso wrote:
> > On Fri, 2013-09-27 at 12:28 -0700, Linus Torvalds wrote:
> > > On Fri, Sep 27, 2013 at 12:00 PM, Waiman Long <Waiman.Long@hp.com> wrote:
> > > >
> > > > On a large NUMA machine, it is entirely possible that a fairly large
> > > > number of threads are queuing up in the ticket spinlock queue to do
> > > > the wakeup operation. In fact, only one will be needed.  This patch
> > > > tries to reduce spinlock contention by doing just that.
> > > >
> > > > A new wakeup field is added to the rwsem structure. This field is
> > > > set on entry to rwsem_wake() and __rwsem_do_wake() to mark that a
> > > > thread is pending to do the wakeup call. It is cleared on exit from
> > > > those functions.
> > > 
> > > Ok, this is *much* simpler than adding the new MCS spinlock, so I'm
> > > wondering what the performance difference between the two are.
> > 
> > Both approaches should be complementary. The idea of optimistic spinning
> > in rwsems is to avoid putting putting the writer on the wait queue -
> > reducing contention and giving a greater chance for the rwsem
> > to get acquired. Waiman's approach is once the blocking actually occurs,
> > and at this point I'm not sure how this will affect writer stealing
> > logic.
> > 
> 
> I agree with the view that the two approaches are complementary to each 
> other.  They address different bottleneck areas in the rwsem. Here're 
> the performance numbers for exim workload compared to a vanilla kernel.
> 
> Waimain's patch:        +2.0%
> Alex+Tim's patchset:    +4.8%
> Waiman+Alex+Tim:        +5.3%

I think I'd like to see a combo series submitted :-)

Thanks,

	Ingo