Figures of merit for time servers

[Eric S. Raymond]

Clark B. Wierda <cbwierda at gmail.com>:
> I think I will try to avoid using the term 'accuracy' in describing
> performance and see how far I can get without it.

Thank you, thank you, thank you!  I thought I was going to have to
invent a performance-evaluation tool, and I am desperately glad
someone else is motivated to tackle the problem.

You wanted a mission - this is yours, if you choose to accept it.  As
always, should you or any of your force be caught or killed, the
Secretary will disavow any knowledge of your actions. This tape
will self-destruct in 5 seconds.

Seriously, here's how this fits in the bigger picture. There are more
features I want to cut to reduce code complexity (a big one is the
KERNEL_PLL code), and refactorings like breaking management of local
clocks out to a separate refclockd.  Increasingly, though, we're going
to face questions about whether these changes degrade performance and
make ntpd unfit for purpose.

It's not going to be enough to merely believe and assert that they don't,
because ntpd's userbase is intensely conservative. We'll need to be 
*demonstrate* it with *evidence*.

Half ot the plan for doing this is the Pi test farm, where we can
cheaply set up a bunch of Statum 1s on identical hardware. The
other half must be the ability to make charts graphing important
figures of merit before and after the changes.
-- 
		<a href="http://www.catb.org/~esr/">Eric S. Raymond</a>