[ntp:questions] Synchronizing Linux clients with Windows Server 2003 NTP

David Woolley david at djwhome.demon.co.uk
Tue Jan 23 07:55:51 UTC 2007

In article <ep3bds$cnl$1 at scrotar.nss.udel.edu>, mills at udel.edu wrote:

> As I said in my message to you, Windows shows dispersion 10s, which is 

Which is a realistic figure for an undisciplined local clock and an 
improvement on the reference implementation.  The reference implementation
makes the assumption, which represents a very small minority of cases these
days, that a server running the local clock driver is being disciplined,
but not by ntpd.  Maybe the reference implementation should add a new
clock driver for this, minority case, and make the local clock driver
report a realistic root dispersion!  (It would make the number one FAQ
for the next few years!  One would need tinker options to disable root
dispersion checks to allow isolated systems to continue to synchronise to
each other.)

> Windows shows a precision of 6, which will cause the ntpd server to 

Which is, I believe, the correct value for the W32Time implementations,
as they only read the time to the clock interrupt frequency resolution.
Most Unix and Unix-like kernels interpolate clock interrupts by either
reading the residue in the counter-time register, or using the TSC 
counter on modern Intel processors, to interpolate.  (The Windows port
of the reference implementation uses the TSC, but, because it is not
supported by kernel code, I believe that the result is not entirely
reliable on a loaded system.)

So, whilst neither value is good, they reasonably accurately represent
the quality of the time that you are actually getting, and the root
dispersion mitigates the excessively low stratum number being used.

More information about the questions mailing list