[ntp:questions] improving ntpd performance on Windows

Dave Hart davehart at gmail.com
Thu Mar 5 11:26:28 UTC 2009


> Unfortunately under W2k the new code does not seem to work as good during
> the first minutes after startup

20000 seconds is 5.5 hours, the length of the roller-coaster ride it
took as I read it.  I see inflection points which I'm guessing are 512
seconds apart where the counter frequency estimate was changed.  I'd
like to see the ntp.log or event log entries for the first few hours
of that w2k run, if available.  I'm particularly interested in the
(frequent) messages showing the ctr frequency observations and running
average.

Meanwhile, I've been whacking on this code more to try to get cleaner
samples to begin with by using a separate timer and callback (though
same thread) for managing the counter frequency, and by starting out
with the history of observations effectively filled with the nominal
performance counter rate, so the adjustment starts happening after one
period (currently testing 512 seconds again after many tests again
with smaller periods), albeit at a moderated rate.  After the first
period, with a depth of 4, the rate used is the average of one
observation and 3 nominal rates.  After two, half from observation and
half nominal.  Those change might affect this behavior.

I'm much more interested in the w2k improvements than vista.  The
blunt object fix of using the native clock is boring and hasn't needed
any fine tuning ;)

Cheers,
Dave Hart




More information about the questions mailing list