[ntp:questions] What is the "best" synchronization possible over the network?

John Ioannidis ntp at tla.org
Fri Feb 13 04:27:22 UTC 2009

Hal Murray wrote:
> In article <49920E17.6050007 at tla.org>,
>  ntp at tla.org (John Ioannidis) writes:
>> The problem setup: two locations, both within the United States, neither 
>> has roof access so no GPS reception is possible.  How do you synchronize 
>> them with better than 50-microsecond accuracy?  Straight NTP over the 
>> Internet doesn't do the trick.  They don't need to actually be 
>> synchronized to "real" time, they only need to be synchronized to each 
>> other.  Assume reasonably unlimited resources (running a private fiber 
>> plant across the continent *is* unreasonable).
>> I've looked into slaving something off a voice-carrier time base, but 
>> that's not accurate enough.  Maybe something over raw SONET would do the 
>> trick?
> SONET gets you frequency, not time.  The tick marks (frames) are not
> synchronized across various SONET links.  Even the frequency isn't
> locked if the two ends come from different phone companies.
> I think you have several basic approaches.
> If cell phones work where you need the time, that may be an alternate
> to GPS.  I don't know what flavor of cell service they use.

Gene Miller suggested these devices:


They are very reasonably priced: under $4K for the full-featured time 
server, plus $3K for the Rb oscillator.  So long as neither VZ nor S go 
out of business, we're in good shape!

> One approach is to get good clocks at both ends and let them coast.
> Atomic clocks are $50K to $100K ballpark.  They have long term
> accuracy of 1E12.  So you get to coast for roughly 50*1E12 microseconds
> or 50E6 seconds.  That's 58 days, so it will be a lot of work to keep
> them synchronized.  I haven't worried about factors of two.  You might
> get better than the spec accuracy if your air conditioning is good...

Actually, 50E6 seconds is 578 days.  It's not out of the question to 
have a third clock always sync-ed to GPS, fed off a 48V battery, and 
drive it to the data centers twice a year to slave the local oscillators.
> Another approach would be to build an extension cord to a place
> where you can get an antenna on the roof, setup a GPS unit there,
> and send the PPS signal to where you need it.  This may require more
> research/testing than you want to do.
> Is roof access really impossible?  How much would it cost?

It involves problems at layer 9 :(

> Do you have any windows?  Have you tried a GPS receiver there?
> With a good timing receiver, you only need 1 satellite.  (You need
> several for setup, only one to keep going.)

No windows, I use Solaris :)  There is, however, cellphone reception in 
the datacenters, so the EndRunTechnologies devices should work.

Thanks again


More information about the questions mailing list