[ovs-dev] [thread 09/15] ofproto-dpif: Lock the expirable list.

[Ben Pfaff]

Ah.

Is it OK for us to keep complaining about you removing it though?

On Tue, Aug 20, 2013 at 01:10:06PM -0700, Ethan Jackson wrote:
> I never removed it, just complained about it.
> 
> Ethan
> 
> On Tue, Aug 20, 2013 at 12:47 PM, Ben Pfaff <blp at nicira.com> wrote:
> > On Sun, Aug 18, 2013 at 02:56:55PM +1000, Simon Horman wrote:
> >> On Thu, Aug 08, 2013 at 02:37:48PM -0700, Ben Pfaff wrote:
> >> > On Thu, Aug 08, 2013 at 02:32:08PM -0700, Ethan Jackson wrote:
> >> > > > Simon presented numbers that showed it to be a valuable optimization
> >> > > > in some cases, otherwise I'd just say get rid of it.
> >> > >
> >> > > If that's the only reason we have it, I vote we ditch it.  It's a new
> >> > > world with multithreading, I'd like to have a clean slate in the
> >> > > direction of thread safety and add optimizations back if they help in
> >> > > this new paradigm
> >> >
> >> > That's the only reason we have it.  If you want to ditch it be my
> >> > guest.
> >>
> >> I accept that multi-threaded ovs-vswtichd is a whole new world
> >> and that dropping optimisations that previously made sense is logical.
> >> And I guess that the best thing is for the situation that lead to
> >> this optimisation to be re-profiled once the multi-threading work is more
> >> complete.
> >>
> >> For reference, my recollection is that this optimisation came
> >> about because it was observed that inserting 100k non-expirable flows
> >> would result in ovs-vswtichd consuming about 100% (of one) CPU running
> >> through the list of flows to see if any of them were ready to be expired
> >> although none of them ever would be. With the optimisation in place
> >> CPU utilisation was reduced to roughly 0% as the list that was traversed
> >> became empty and the system was otherwise idle.
> >
> > I think it's perfectly reasonable to reintroduce the optimization.