[ntp:questions] NTP Performance on WINNT

David Woolley david at ex.djwhome.demon.co.uk.invalid
Sat Jul 25 16:58:30 UTC 2009

paul wrote:

> Hi David, I am wondering is there any other means to profile NTP
> performance, like some kind of hardware setup to measure time offset
> of two machine?

You need access to the source code of the clock interrupt service 
routine and the ability to directly right to something like a parallel 
port if you want to measure the true offset (relative to something like 
GPS pulse per second).  The former, at least, you are not going to get 
under a closed source OS, like WINNT.

Also remember that Windows never gives the time to application programs 
to better than 1ms, and this can degrade to more like 16ms.  Also 
Windows offers no guarantees of scheduling latencies between real 
external events and when application code runs to service them.

Another option might be to install some sort of clock card that can be 
read through a low latency device driver, but then you would be better 
using that synchronised clock card, rather than the software clock, for 
your timing.

Although you might be able to drive a real (non-USB) parallel port, from 
application code, with fairly low latency the results would only be 
meaningful for a very unloaded machine, as, on a loaded machine, you 
wouldn't really know where you where in the system tick interval, when 
you read the software time.

I'd suggest, therefore, that you will need to write a device driver, for 
good accuracy, and be Microsoft or their subcontractor, for full accuracy.

More information about the questions mailing list