[ntp:questions] help with setting up NTP on windows with a USB GPS

Dave Hart davehart at gmail.com
Sun Dec 6 01:26:17 UTC 2009


On Dec 6, 12:48 am, "Richard B. Gilbert" wrote:
> My NTP server is an old, old, Sun Ultra 10 workstation (I'm guessing)
> ca. 1998.  It gets the job done.  If there were any 486/33 processors
> left, one of those fifteen or twenty year old machines could do a very
> good job of running NTPD.

For the purposes of ntpd, the older the hardware the better in many
ways.  As performance keeps being goosed out of various nooks and
crannies of hardware and software underneath ntpd, consistency is
often lost in favor of throughput.

 o Older processors and associated chipsets were simpler designs with
extremely predictable and repeatable performance.  Everything in the
x86 space doing hyperthreading or multiple cores has lost in terms of
repeatable timing.
 o As is often mentioned, gigabit ethernet has much higher jitter out
of the box, and due to switches, even with tuning, as compared to
10/100.
 o Traditional RS-232 serial ports not connected via USB are becoming
rare, making it more difficult to find a suitable PPS hardware
interface.
 o When it comes to Windows, Vista and later essentially disable
ntpd's Windows-only interpolation hack by increasing the system clock
precision from 2^-6 to 2^-10 or 2^-11.  With the scheduler running at
millisecond (~= 2^-10) granularity, ntpd simply can't sample the clock
and cycle counter fast enough to interpolate.  This means ntpd's
precision drops from around 2^-20 to 2^-10 or 2^-11, or microsecond to
millisecond.  Finding hardware that supports XP/2003 and older gets
more difficult every day.  The situation is probably better on servers
where use of timeBeginPeriod by _any_ application can be avoided,
leaving the system clock precision around 10-15ms.  I haven't tested
servers newer than 2003 to know for sure.

In short, ntpd likes speed, but it likes consistency even more.

Cheers,
Dave Hart




More information about the questions mailing list