[ntp:questions] comparing 2 NTP implementations

David Woolley david at ex.djwhome.demon.co.uk.invalid
Sun Jan 27 09:18:32 UTC 2008


Folkert van Heusden wrote:

> I would like to compare 2 NTP implementations. What would be the best
> way?

The biggest problem is finding out the time on the machines without 
using NTP.  One approach is to use a simulator, but that assumes that 
the simulator correctly represents clock imperfections and changes in in 
the environment.

It is also fairly easy to output a quite accurate indication of the time 
that the machine thinks it has, provided that you have a local (non-USB) 
parallel port.  However, the problem there is that they will not output 
at the same time, which means that you cannot use very simple hardware 
to measure the difference, but will need hardware that accurately log 
both the actual time of the report and the time the reporter thought it 
had.  The clock for this probably doesn't have to be too accurate, 
providing that you monitor your source of true time frequently, but it 
does have to have good precision and predictable latency.

You cannot output at the same time because of indeterminate interrupt 
latency and because modern systems interpolate between clock interrupts 
and correct the time by adjusting the interpolation, not by aligning the 
clock interrupts onto the exact 100ms.

The other requirement, that has been noted in the recent chrony 
discussion, is that you must run the test with real workloads on the 
machines.

> I was thinking of configuring 7 upstream servers on these 2 physical
> servers and then on a third pc (which is also synced against these 7)
> check the difference?

You need to have physically distinct servers, so that clock variations 
are not correlated.  They ought to be in different environments. 
Alternatively, the server needs to directly read a high stability 
hardware clock and simulate perturbations.

> 




More information about the questions mailing list