[time-nuts] PC clock comparison software?

Hal Murray hmurray at megapathdsl.net
Wed Dec 20 05:59:31 EST 2006

> For test purposes I've occassionally set the time server polling
> routine to once a day and seen 20 minute errors on delinquent PCs.
> Other PCs might stay within a minute.

1 min per day is 694 ppm
20 min per day is 13880 ppm

ntpd calls the crystal inaccuracy drift.  ntpd can't handle a clock drift of 
more than 500 ppm.  100 ppm is reasonable.  That's 8.64 seconds per day.  
(This PC is 115.755 ppm)

Long before serious time keeping software, we used to set the equivalent of 
the drift correction by hand.  That got us to a second per week or so on 
machines that were in a server room with reasonable air conditioning.

My guess on your systems that are off by 20 min would be interrupts 
disabled/busy for long enough so that clock interrupts get lost.  It's easy 
to tickle that if you do a lot of disk activity on a Linux box with DMA 
disabled on the disks.  (Or was a few years ago.  I haven't tried it 
recently.)  Were all those way-off clocks running slow?

If you can get ntp running on all/most of your systems, it's pretty easy to 
setup one system to monitor all the others.  Turn on logging (peerstats) and 
add each system you want to monitor as a server (in your config file) with 
the noselect option.  Tweak minpoll/maxpoll if you want to adjust the amount 
of data you collect.

> Our network is about 30 PCs on fiber OC12 ( 622Mbps) ATM with 100mbps
> ethernet network branches. In the near future some video conferencing
> and remote telescope users will add to the chunkiness of the network
> variable load.

I wouldn't expect network troubles unless parts of your network are seriously 

It's pretty easy to see busy DSL links.

The ntp log files will tell you the round trip time as well as the offset of 
the other system.  ntpd assumes network delays are symmetric.  If you know 
the clocks on both ends are good (say both have GPS) you can measure the 
network delays in each direction.  That is the observed offset is really an 
asymmetric delay.

These are my opinions, not necessarily my employer's.  I hate spam.

More information about the time-nuts mailing list