[ntp:questions] ntp server specs

unruh unruh at wormhole.physics.ubc.ca
Thu Apr 21 00:51:00 UTC 2011


On 2011-04-21, Richard B. Gilbert <rgilbert88 at comcast.net> wrote:
> On 4/20/2011 5:49 PM, unruh wrote:
>> On 2011-04-20, C BlacK<rblak at non.net>  wrote:
>>> If I am setting up an ntp server for a lot machines, I have some questions.
>>>
>>> If the timing on client machines needs to be tight, should the ntp server be an x86 machine with a less than accurate clock
>>> or a ntp server with a very accurate clock.
>>
>> ? Clearly the server should have as good a clock as possible. And it
>> should not CPU speed type power management. And it should not be on a GB
>> network.
>> Note that if you run the server with a GPS PPS you can get it down to
>> 5usec accuracy, at which point the type of machine is largely
>> irrelevant.
>>
>>>
>>> I want the clients to track closely to server.
>>
>> What does "track closely" mean? 1 hour? 1 sec? 1ms? 1us, 1ns?
>>
><snip>
>
> It's reasonable to expect NTPD to track the correct time within 10 
> milliseconds or better using a server that is close to you in *Net 

More like 10-100us.

> Space*.  It's conceivable that a server could be physically no more than 
> five miles away and yet have a network path between the two with a
> link that is 200 miles long.  Always remember that the shortest network 
> path will usually give you the best results.
>
> In principle, you could be in New York City and use a server in Tokyo. 
> In practice the "phase noise" will kill you!  There are hundreds or even 
> thousands of possible paths between New York and Tokyo and you could get 
> a different one with each request and thus a different delay with every 
> request!!!!!

IF the path were the same coming and going this would not matter. That
is of course a big IF.

>
> If you have a budget that will stand it, you can get a hardware 
> reference clock such as a GPS Receiver, LORAN receiver, etc, etc.
> A good reference clock should be able to keep you within 50 nanoseconds 
> or better.  Keep in mind that, while the PPS output of a GPS or other 
> timing receiver can be within 50 nanoseconds of true time, getting the 
> signal into a computer can introduce delays that are extremely difficult 
> to measure!

Yes, they are at the usec level. Worse if the interrupt handler routine
does not itself do the time-stamping of the interrupt. 
So, no, it cannot keep you within 50ns. But it can within about 1usec (
with ntpd, more like 5-10usec)

>
>




More information about the questions mailing list