Performance of 4.x vs 5.x (Re: Lifetime of FreeBSD branches)

[Jon Dama]

It's different, yes.  But the trouble is that you need a controlled
interrupt source--i.e., you have to have some concept of when an "event"
might have been handled (were it not for such and such activity).

I posit that without that counterfactual talking about PREEMPTION is
meaningless.

The technique I mentioned--measuring and comparing the jitter was intended
to quash measuring the performance of the network stack itself.

Do you have an idea how you can pose that counterfactual in a synthetic
arrangement more closely connected with the problem at hand?

...

-Jon

On Wed, 25 May 2005, Kris Kennaway wrote:

> On Wed, May 25, 2005 at 03:33:39PM -0700, Jon Dama wrote:
> > Could this be quantified by setting up a synthetic experiement:
> >
> > 1) one machine uses dummynet to generate a uniform packet/sec stream
> > 2) another machine has a process receiving those packets and recording
> >    their arrival relative to the local TSC.  afaik, the TSC is the only
> >    source of wall-time that doesn't involve a system call.  Is that right?
> >    Are the TSCs synchronized on SMP systems?
> > 3) Generate another source of activity on the receiving machine to
> >    estimate the effect of PREEMPTION relative to the (lack of) quiescence.
> > 4) use the jitter in the TSC deltas to infer the effect of preemption
>
> That would be attempting to benchmark something entirely different.
>
> Kris
>