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

Terje Mathisen terje.mathisen at hda.hydro.com
Thu Apr 13 13:06:27 UTC 2006

Ry wrote:
> Here's the actual link:
> http://research.microsoft.com/users/misard/papers/aes2004.pdf
> Basically, the MS guys moved most everything into a kernel mode driver
> and alter NTP to poll every 4 seconds. No problem for their
> application, since all devices are on the same LAN as the master audio
> clock. I can't decide if the ±50µs accuracy is impressive or not...
> but obviously they wouldn't have modified NTP so heavily if it would
> have met the application's needs out of the box.

They have solved two core problems:

1) The Microsoft kernel clock has approx 10 ms resolution, which is far 
too coarse.

2) Packet processing latency variations impact short-term stability.

They also claim to have invented 'burst' mode, i.e. they send 5 packets 
during each polling period and select the median. :-)

If the kernel timestamping and improved OS clock resolution is or could 
become available in a standalone driver, then this would represent a big 
boost for Windows-based NTP servers.

In fact, they could eventually provide comparable performance to FreeBSD 
running on the same HW! :-)


- <Terje.Mathisen at hda.hydro.com>
"almost all programming can be viewed as an exercise in caching"

More information about the questions mailing list