[ntp:questions] Re: NTP client on Windows platform provides less accurate results then on the UNIX or Linux. Why?

Ry malayter at gmail.com
Mon Apr 3 04:47:35 UTC 2006


On 4/2/06, Danny Mayer <mayer at ntp.isc.org> wrote:
> That should have no more affect on NTP than anything else running on the
> box. We have seen no such affects. If you want to report a problem then
> you need to send much more detail.

I'm not reporting a specific issue, I'm just pointing out potential
problem areas for the original poster to look at. Additional,
systemic, asymmetric latency on the Windows machines would certainly
make them less accurate than his Linux boxes. I am confident, for
instance, that there is non-trivial latency added by Symantec Client
Security running on Windows when it tries to match IDS and AV patterns
on all network activity. This is something Linux boxes wouldn't do.
I've never tried to measure it, but it would be interesting to see
what it does to NTP performance.

> I'm really not sure why you think that restarts have such a great effect
> on NTP. It's designed to converge quickly and certainly no more than
> about a day.

On both of my stratum-2 boxes on Windows 2003sp1, they converge to
withing a half-dozen milliseconds in about 8 hours. But after running
for several days, they seem to converge to within ~1 ms offset. I
don't know why that is, but NTP certainly does seem to get better on
my Windows machines over longer periods of time. When I frequently
restart a machine, I never get much better than 10ms accuracy. I don't
know why that is, other than NTP isn't disciplining the clock during
the actual shutdown and restart.

Of course drift would only be about 2 ms per minute of NTPd downtime
on a 30ppm machine. But it seems to me after NTPd is down it over
corrects a bit with the drift calculation and I get a few swings
before it settles in after 8 or so hours. I posted a chart of this
"restart over correction" on an XP box at:
http://www.malayter.com/NTPovercorrectAfterRestart.gif
Each of the dips followed by an over correction (including the huge
over correction in the middle that I can't explain) corresponds to a
reboot of XP SP2. Ntp.drift is currently 9.208 on the system. Four
low-jitter stratum-2 servers near me are the time sources (they're run
by university CS departments and even a national laboratory, so I
think they're pretty good).

Please understand that all I am trying to do is help the original
poster figure out why his XP workstations aren't performing as well as
his Linux boxes. I am not "complaining" about anything or assaulting
your work in any way. I think the Meinberg package of NTPd is
fantastic. My basic concern was that the original poster might be
comparing the NTP accuracy of Linux machines that are always on to
Windows XP workstations that are frequently restarted or even shut
down for periods of time.

Regards,

  Ryan




More information about the questions mailing list