[ntp:questions] Re: Setting Up NTP Subnet

John Pettitt jpp at cloudview.com
Wed Mar 8 21:20:07 UTC 2006

Bill Myers wrote:
> On 3/7/06, Richard B. Gilbert <rgilbert88 at comcast.net> wrote:
>> Bill Myers wrote:
>>>  I have questions regarding best practices on architecture of NTP
>> subnets.
>>> I've thoroughly read "Notes on Setting up a NTP subnet" from the NTP
>> sites.
>>> Is there a need for multiple external time source TECHNOLOGIES.  That
>> is, is
>>> it sufficient to use multiple (8) GPS receivers with ACTS dial backup
>>> and Rubidium clocks?   This is a sizeable investment for the
>> infrastructure,
>>> but is it the right way?  Is it necessary or desirable to have other
>> sources
>>> such as radio and/or Internet?
>>> In case you have not picked up on it, I'm talking about an appliance
>> here.
>>> I find it odd that one cannot NTP peer these appliances.  This suggests
>> to
>>> me that we should create a stratum 2 tier peer layer a stratum 3 tier
>> that
>>> peers and serves time to the endpoints.
>>> More dysfunction?  The plan is for the stratum 2 tier, which is also the
>>> time distribution tier, to be the cache DNS for some high-volume data
>> center
>>> environments.
>>> _______________________________________________
>>> questions mailing list
>>> questions at lists.ntp.isc.org
>>> https://lists.ntp.isc.org/mailman/listinfo/questions
>> It depends strongly on your needs for correct time and reliability.  You
>> have not really specified your requirements. Are you serving a single
>> site or multiple sites?  How far can your time deviate from UTC?   How
>> much money are your willing and able to spend?
>> If you absolutely must have sub millisecond accuracy and 100 percent
>> reliability then you need to allow for:
>> 1. Failure of a reference clock (GPS receiver)
>> 2. Failure of the GPS System (insane satellites, jamming of GPS
>> frequencies, etc.)
>> 3. Failure of one or more of your lowest stratum servers.
>> 4. Things neither you nor I have thought of or can think of!
>> So tell us more and maybe we can give you better advice.
> Reliability and synchronization is most important.  <10 millisecond accuracy
> is probably sufficient.  Large financial with 100's of locations.
> International.  Has experienced financial loss due to NTP failure.
> ~Bill

There are a few issues to consider

1) Do you need internal consistency or internal consistency and external accuracy?   This will drive
your need for ref clocks.

2) With 100's of location can a location continue to operate as an island due to partial network
failure.  If so how much holdover time do you need for accurate timing? If for example a location
could operate on an external Internet connection while the internal net is down you may need to
consider having ref clocks at each site.

3) Do applications need to be aware of the current state of timekeeping?  It's one thing to specify
an error limit of 10ms but how will you know if you've exceeded it and what should happen?  Should
the applications refuse to accept mis-timed transactions?  What happens when an island rejoins the
net but has not yet converged on network time?


More information about the questions mailing list