[ntp:questions] question regarding NTP configuration for clusters, and "cluster time" stability

Unruh unruh-spam at physics.ubc.ca
Tue Sep 29 23:12:58 UTC 2009


"rotordyn at yahoo.com" <rotordyn1 at gmail.com> writes:

>On Sep 29, 4:12=A0pm, Unruh <unruh-s... at physics.ubc.ca> wrote:

>> Oddly enough, making sure that all of the computers are synchronized to
>> UTC is probably the best way to ensure that they are all synchronized to
>> each other.

>Perhaps, but I haven't sorted out a way to achieve that with the
>reference
>NTP implementation, at least in a configuration that has redundancy
>built-in. Picking one node to be the root of the internal ntp
>heirarchy is
>straightforward enough, but then I have to manage failover myself.

Well, have 4 of them be the internal top level, and have the rest of the
nodes use all 4 as their servers. Then make sure that if one looses
sync with UTC, it drops down in stratum so that the rest of the machines
will ignore it. 
 

>I can do that too, but I run into chicken and egg issues, in that the
>cluster software needs the nodes to be in synch before starting,
>so NTP must already be running.

Uh, yes, the computer must run in  order for the computer clock to run.
That is Always true. 


>> For example, you could run a PPS line to each computer's
>> parallel port and use the interrupt from that to sync that computer to
>> that PPS. Or you could sync a bunch =A0of them to their own PPS source (
>> eg a number of gps).

>With thousands of deployed systems, adding hardware simply
>isn't an option.



>> If the peer looses it external utc reference you could always have it up
>> its level, so the rest would stop using it as a reference.

>That's an interesting idea. Is there a way to implement that as a
>configuration
>option, or is that something I'd have to code myself? (The latter is
>ok if necessary)

>> One of the advantages of GPS is that it too exists in the field. And on
>> house tops, and in the canyons ( well at least rooftops) of the concrete
>> jungle.

>Besides the issue of thousands of systems in use, there's the added
>one that we're always deployed under a roof, never on top of it. :)

Well, yes, but then lines run through the roof or a window are usually
pretty easy. 

>thanks,
>tim





More information about the questions mailing list