[ntp:questions] Tighter regulation?

David Woolley david at ex.djwhome.demon.invalid
Fri May 24 09:39:01 UTC 2013


Mischanko, Edward T wrote:
> 
>>
>>> On 2013-05-21, Mischanko, Edward T <Edward.Mischanko at arcelormittal.com>
>> wrote:
>>>> My concern is that too much data is being thrown away when polling
>>>> above 256 seconds and that allows excessive wandering of my clock.
>> If too much data is being thrown away, it would be because the poll
>> adjust algorithm has chosen too high a poll interval (or you have too
>> high a minpoll) for the noise statistics.
> [Mischanko, Edward T] 
> 
> This is exactly what I am saying!!  This needs to be corrected; I consider it a bug.

How would you modify the algorithm that chooses the poll interval. 
Remember that getting too low a poll interval will make ntpd track short 
term variations in network latency and prevent it getting an accurate 
estimate of the crystal frequency error.

Please provide a detailed specification of your proposed algorithm.

> Too much assumption is made that everyone will have the perfect computer and 
> the perfect network when configuring these various filters.  What works on

A lot of allowances have been made for the real world.  The only 
allowance that has not been made is that it assumes that crystal 
frequency errors are essentially random, and thus can get into 
difficulty if a (normally very cheap) motherboard crystal is in a 
thermal environment that exhibits non-gaussian behaviour.

> the blackboard does not always work in reality.  My computer has a
> -19 precision but it can't keep time inside 1 millisecond with default
> Settings; go figure.

I figure that you don't understand the various latencies that exists in 
the system and network, and the difficulties of interpolating clock 
ticks on Windows.  The precision figure, for Windows, is almost 
certainly optimistic because of the hacks ntpd has to go through to 
interpolate the clock in user space.

The reason for the difficulty in interpolating is that Windows does not 
achieve anything like the timing resolution in the Windows API 
architecture.  Earlier versions of Windows NT could only supply time to 
applications to a resolution of 16ms (before ntpd started doing user 
space interpolation, NT 3.5 would return an ntpd precision of -6). 
Intermediate ones had a variable resolution, depending on the use of 
multi-media timers, but I don't think it was better than 1ms.  I'm not 
sure of the exact parameters to the latest generation; but I don't think 
ordinary applications will be seeing anything close to 2 microsecond 
resolution.  (Changes of multimedia timer rates, causes upsets.) 
Basically, you may find that the main error component in the time 
supplied to ordiary Windows applications is due to the limitations of 
the Windows time reading APIs.

Also, I don't remember your ever describing a test rig that is capable 
of comparing the system time on your servers to true time.  Offset is 
not offset from true time.  Without special instrumentation, or an 
accurate theoretical model, you cannot assume that high NTP offsets 
represent errors in the machine time keeping.
>



More information about the questions mailing list