[ntp:questions] frequency adjusting only

Unruh unruh-spam at physics.ubc.ca
Tue Apr 22 15:53:05 UTC 2008

m.louvel at gmail.com (maxime louvel) writes:


>I am actually using 4 public NTP server, which whom my node-1 synchronises
>Then node-1 broadcasts the time to all the other nodes in the subnet.

>My goal is to achieve a synchronisation between the nodes (not with the
>public NTP server) within 50 usec.

That should be possible even with ntp. chrony will do a somewhat better
job. On a "close" net like yours, the problem with ntp is that it does a
poor job of tracking the frequency changes that come with the system's temp
fluctuations because of doing variable amounts of work. chrony does a
better job of that. But 50us should be possible with either. 

>I don't care if the synchronisation to the public NTP is accurate around
>half a second.
>I think it's possible, because the nodes are close to each other and every
>communication is using gigabit ethernet (cards + switch).

It is the latency not the speed of the switches that is important. 

>I achieve to get a small offset when nothing else than linux and ntpd is
>running on the nodes
>(offset arount 10 usec).

>I have a program which sends data from one node to another.
>The sending/reception are timestamped and the delay is then computed.

>I would like the measured delay to be constrained in a 50usec wide range.

YOu have NO control over the delay. That is in the switches. 

>Do you think it should be possible ?

>I have tried to run the test for 24h and didn't get the expected results.
>The range of the measured delay is 2 milliseconds (between -1ms and +1ms)
>with some spikes (ten or so for a milion measurments). The mean delay is
>-0.066087 (ms) and the std dev is 0.22849
>Do you think the spikes can influence ntp  ?

It is designed to get rid of spikes.
What in the world do you have on that net? That is bad behaviour on the
part of the switches. Are they really that overloaded?

>thanks very much


>On Tue, Apr 22, 2008 at 1:15 PM, Danny Mayer <mayer at ntp.isc.org> wrote:

>> maxime louvel wrote:
>> > Thanks for your answer,
>> >
>> > I can't have a step on the clock because that would screw up my
>> > applications.
>> > However if I keep the load within a certain range I should fine, don't I
>> > ?
>> > I am synchronising one node to several public NTP servers, and the
>> > others
>> > nodes are synchronised to the first one.
>> > There are 2 to 24 nodes in my sub net.
>> > Do you think that should be feasible ?
>> >
>> > Maxime
>> >
>> You need at least 3 and preferably 4 public NTP servers in order for your
>> local server to make decisions about which one is giving the most accurate
>> and reliable time at any moment. 2 is not enough since there's no way to
>> decide which is better. 3 allows 2 to gang up on the third.
>> 2-24 nodes is nothing. Even bombarding the NTP server with queries will
>> barely be noticed. Don't forget at a minimum a node won't send a query more
>> frequently than once every 64 seconds unless you are using a badly
>> implemented NTP client.
>> Danny

>Maxime Louvel
>0044 7964 5555 80
>43 Allen road
>Whitemore reans
>WV60AW Wolverhampton
>United Kingdom

More information about the questions mailing list