svn commit: r304436 - in head: . sys/netinet

[Slawa Olhovchenkov]

On Fri, Aug 26, 2016 at 02:32:00PM -0700, Adrian Chadd wrote:

> Hi,
> 
> It's pcb lock contention.

Not sure: only 5% of all time.
And same 5% for tcbhashsize = 65K and 256K.
Or you talk about some more thin effect?

> 
> On 26 August 2016 at 08:13, Slawa Olhovchenkov <slw at zxy.spb.ru> wrote:
> > On Fri, Aug 26, 2016 at 04:01:14PM +0100, Bruce Simpson wrote:
> >
> >> Slawa,
> >>
> >> I'm afraid this may be a bit of a non-sequitur. Sorry.. I seem to be
> >> missing something. As I understand it this thread is about Ryan's change
> >> to netinet for broadcast.
> >>
> >> On 26/08/16 15:49, Slawa Olhovchenkov wrote:
> >> > On Sun, Aug 21, 2016 at 03:04:00AM +0300, Slawa Olhovchenkov wrote:
> >> >> On Sun, Aug 21, 2016 at 12:25:46AM +0100, Bruce Simpson wrote:
> >> >>> Whilst I agree with your concerns about multipoint, I support the
> >> >>> motivation behind Ryan's original change: optimize the common case.
> >> >>
> >> >> Oh, common case...
> >> >> I am have pmc profiling for TCP output and see on this SVG picture and
> >> >> don't find any simple way.
> >> >> You want to watch too?
> >> >
> >> > At time peak network traffic (more then 25K connections, about 20Gbit
> >> > total traffic) half of cores fully utilised by network stack.
> >> >
> >> > This is flamegraph from one core: http://zxy.spb.ru/cpu10.svg
> >> > This is same, but stack cut of at ixgbe_rxeof for more unified
> >> > tcp/ip stack view http://zxy.spb.ru/cpu10u.svg
> >> ...
> >>
> >> I appreciate that you've taken the time to post a flamegraph (a
> >> fashionable visualization) of relative performance in the FreeBSD
> >> networking stack.
> >>
> >> Sadly, I am mostly out of my depth for looking at stack wide performance
> >> for the moment; for the things I look at involving FreeBSD at work just
> >> at the moment, I would not generally go down there except for specific
> >> performance issues (e.g. with IEEE 1588).
> >>
> >> It sounds as though perhaps you should raise a wider discussion about
> >> your results on -net. I would caution you however that the Function
> >> Boundary Trace (FBT) provider for DTrace can introduce a fair amount of
> >> noise to the raw performance data because of the trap mechanism it uses.
> >> This ruled it out for one of my own studies requiring packet-level accuracy.
> >>
> >> Whilst raw pmc(4) profiles may require more post-processing, they will
> >> provide less equivocal data (and a better fix) on the hot path, due also
> >> to being sampled effectively on a PMC interrupt (a gather stage- poll
> >> core+uncore MSRs), not purely a software timer interrupt.
> >
> > Thanks for answer, I am now try to start discussion on -net.