[ntp:questions] NTPD can take 10 hours to achieve stability
unruh at wormhole.physics.ubc.ca
Wed Apr 20 21:05:46 UTC 2011
On 2011-04-20, David Lord <snews at lordynet.org> wrote:
> unruh wrote:
>> On 2011-04-20, David Lord <snews at lordynet.org> wrote:
>>> unruh wrote:
>>>> On 2011-04-19, David Lord <snews at lordynet.org> wrote:
>>>>> unruh wrote:
>>>>>> On 2011-04-19, David Lord <snews at lordynet.org> wrote:
>>>>>>> When I first compared chrony with ntpd there was no contest
>>>>>>> but more recent experiments with chrony had periods of
>>>>>>> severe instability much worse than ntpd.
>>>>>> More recent means what? What version of chrony? (there was a bug found
>>>>>> in the past two weeks that was introduces a few months ago which did
>>>>>> result in instability)
>>>>> More recent than November 2009.
>>>>> Probably between Dec 2009 and Jan 2010 and p4x2400c.
>>>>> chrony 1.23
>>>> OK well before that change.
>>>>>>> P4X2666 with chrony
>>>>>> Not at all sure what I am supposed to see. I have no idea what the graph
>>>>>> axes represent? What is 1.1k
>>>>> X-axis is in hours
>>>>> Y-axis is offset in us so 1.1k = 1100us
>>>> WOW. What kind of network are you attached to? Even on an ADSL link
>>>> through the phone company, I
>>>> was getting in the tens of usec (not ms) as the offsets of chrony.
>>>> (checked by a gps receiver attached to the local computer).
>>> I'm on ADSL-1 with 2 Mbit/s down and 288 kbit/s up. Latency
>>> is what BT delivers and has been as low as 12ms up to 50ms or
>>> more but mostly it's about 18ms to nearest sites (when exchange
>>> gets near capacity the latency jumps up by 15ms or more same as
>>> when interleave is enabled on ADSSL-2).
>>> Although the peaks of the graph are at about 1ms the stats
>>> have "System time" mostly being in the 10s of usec.
>> Not sure what you mean by that. What I mean is , what are the measured
>> offsets? (Of course if you have a GPS PPS that you could use as a
>> reference-- not a source of time-- that would make much clearer what is
>> going on.)
> My logs have eg:
> Reference ID : me6000g
> Stratum : 2
> Ref time (UTC) : Sat Jan 9 07:12:02 2010
> System time : 0.000001 seconds slow of NTP time
OK, that is the kind of thing chrony should do. Ie, to discipline the
system clock to usec levels
> Frequency : 0.125 ppm fast
That is a really small drift rate. Just lucky I guess.
> Residual freq : -0.010 ppm
> Skew : 0.223 ppm
So the frequency is really pretty stable.
> Root delay : 0.001297 seconds
> Root dispersion : 0.007492 seconds
That is a huge root delay and dispersion. Where are your servers?
> I've not yet checked all of the systems that were tried
> with chrony but these servers are all in use so it's not
> easy to just install chrony and expect it to work.
Sure it is. Just install it, configure it and start it ( switching off
ntpd first). Of course if your server is supposed to keep really good
time all the time, then this might cause some initial problems as you
>>>> And what is it about the graph that makes you believe you are seeing
>>> There is nothing in that session but later graphs now
>>> wrapped (as in loops around after a year) showed much
>>> worse offset variations than the systems using ntpd at
>>> which point I went back to using ntpd.
>> Certainly not my experience. And the lack of evidence makes it hard to
>> fix any problems if they are there.
More information about the questions