[ntp:hackers] Server timestamps going backwards by up to 2 milliseconds

David Mills mills at udel.edu
Sat Dec 26 18:34:48 UTC 2009


Terje,

I get my head on crooked in these things myself, but a nontechnical 
argument may be that, since the ordinary clock reading lops off a 
certain number of bits in each reading, the apparent time is always less 
than or equal to the true time. The fuzz adds a random value so that the 
average apparent time is equal to the average of the true time.

Don't make be say that again; I will probably get it wrong.

Dave

Terje Mathisen wrote:

> Kurt Roeckx wrote:
>
>> On Sat, Dec 26, 2009 at 03:39:33PM +0100, Terje Mathisen wrote:
>>
>>> Stock NTP (no fuzz or other tweaks) would then move said OS clock
>>> backwards by about p/2, making the error go from -p/2 to p/2 over
>>> the length of a tick, and making the average error p/4.
>>
>>
>> If you got T and want to shift the error to get a smaller average
>> error you need to add p/2, not substract it.  The error would now
>> be Tx - (T+p/2) and be between -p/2 and p/2 with an average of 0.
>
>
> We do agree, but my description was unclear:
>
> By adding fuzz to the OS time stamps, NTP will make each tick happen 
> sooner. I looked upon that as moving backwards in time, even though 
> the clock has to go faster for a while to get to this position.
>
> Sorry!
>
> Terje
>



More information about the hackers mailing list