[ntp:questions] NTP - best practice if there is a local stratum 2 server

Richard B. Gilbert rgilbert88 at comcast.net
Sun Jun 28 18:14:07 UTC 2009

Dave wrote:
> I'm trying to set up the time on a couple of machines at the University 
> of Washington and have tried two ways:
> 1) Sync both to the ntp pool - not very successful, as they kept 
> differing by enough to cause 'make' to say files were created in the 
> future by a few ms - no doubt due the NFS exported drives.
> 2) Set one via ntpd (call is master) from the ntp pool, then the other 
> local machine (slave) read the time from the master. That appears to 
> work quite well. They at least agree on the time pretty closely now. The 
> absolute accuracy might not be as good as possible with a local time 
> server, but for practical purposes it works well.
> Searching around, I found that the university does in fact have their 
> own strtatum 2 time server - I assume in another department to where the 
> machines I'm using are located. The server is
> bigben.cac.washington.edu
> I assume, for absolute accuracy (which is not important here, but I'm 
> intersted), this is going to be better than any attempt to set the time 
> from the ntp pool. Is that correct?

Accuracy should be in the very low milliseconds; e.g. plus or minus five 
milliseconds.  You should configure a minimum of four timeservers using 
systems near you in "net space".

A system may be physically within a few hundred yards and still be 
several miles away as far as signal transmission is concerned.
> But setting the time from the local server is also going to be cause a 
> problem if that server fails for some reason.
That's one of the reasons why you are advised to configure four or more 

> Is there any way to make use of a local server if it's working, and then 
> use the pool if its not?

Configure four servers and you can survive the loss of one.  Configure 
five and you can survive the failure of two.  Configure seven and 
survive the failure of three.  Failure can mean either incorrect time or 
  just not responding to queries.
> To be honest, this is all a bit academic, as there is no real practical 
> need to have an accuracy of better than a minute or two on these 
> machines. In fact, an hour wrong would not make much difference but it 
> might irritate a few people.
> But it is important the two machine agree with each other, as otherwise 
> files appear to be created in the future by a few ms, which screws up 
> the 'make' program some times.

The machines should agree with each other reasonably well; e.g. within a 
  hundred or two hundred microseconds.  You MAY get better than that but 
remember that keeping time was not, and will not be, very high priority 
in the design of any computer system.  With some you may consider 
yourself lucky if the date is correct.

It's easier to get and keep close synchronization if you have have some 
sort of a "hardware reference clock" such as a GPS timing receiver, a 
WWV receiver, etc.  GPS timing receivers are available from $100 US and 
up.  The bottom of the scale is a bare circuit board.  Moving upscale 
you can get PCI cards, stylish tabletop devices, rack-mount equipment, etc.

For about $1200 US you can get a clock on a PCI card that will plug into 
your computer.  It will have an oven controlled crystal oscillator and 
provision for a GPS input.  Talk to Symmetricomm and/or Meinberg 
Funkurhren; both have such hardware available.

More information about the questions mailing list