Charly's Column – ntpd

Time is the topic I'm focusing on right now. On the one hand, I am of course happy to be celebrating the 25th anniversary of Linux Magazine and the fact that I have been allowed to contribute this column for more than 15 years. On the other hand, I am currently working on configuring my time servers. In the last issue [1], I briefly touched on the topic when we looked at ntpviz, the statistics visualization tool for the Network Time Protocol daemon (ntpd).

How to tune the time server for maximum accuracy using quite simple means was the topic in the October 2017 issue [2]. (Is it really already two years ago? How time flies.) At the time, I picked up a high-precision pulse per second (PPS) signal from GPS satellites. This allowed me to line up ticks received locally or from remote time servers to achieve microsecond accuracy on my server. Apart from scientific applications, nobody really needs that, but it's cool, so it was done.

When I tackled this configuration back in 2017, it was still by a fairly circuitous route. Feeding the PPS signal to ntpd required additional software, which I found on GitHub. It's easier today. The two daemons involved, gpsd and ntpd, work hand in hand on my Debian 10 – and without any extra software.

I only have to tell the GPS daemon in the /etc/default/gpsd configuration file where to find the GPS hardware. In my case, it is connected to a USB port, so I need the following line:

DEVICES="/dev/ttyUSB0"

Also the ntpd configuration file is quickly completed with Listing 1. Lines 1 to 4 probably already exist in the file. It is important that one of them ends with the prefer keyword. Lines 6 and 7 include the GPS time signal, but only for statistics. The noselect keyword prevents it from actually being included in the calculations, because it is not very accurate. Lines 9 to 11, on the other hand, provide accuracy and bridge the gap to the PPS signal, which announces the beginning of a new second with high precision.

01 server 0.de.pool.ntp.org prefer
02 server 1.de.pool.ntp.org iburst
03 server 2.de.pool.ntp.org iburst
04 server 3.de.pool.ntp.org iburst
05
06 server 127.127.28.0 minpoll4 noselect
07 fudge 127.127.28.0 refid GPS
08
09 server 127.127.22.0 minpoll 4 maxpoll 4
10 fudge 127.127.22.0 flag3 1 refid PPS
11 tos mindist 0.2

A Little Crazy

This allows my time signal to fluctuate by only a few millionths of a second (Figure 1). By the way, the deflections in the right part of the graph were caused by me frivolously moving the GPS antenna on the window sill.

Figure 1: The computer time deviates from the ideal by just +/-2 microseconds. The irregularity starting at eight o'clock is a human-shifted-antenna fault.

Of course, my experiments did not go as smoothly as the story thus far suggests. In the first attempt, my time signal was no more accurate than before, whatever I did. After tearing my hair out for a couple of hours, it turned out that you can buy some GPS hardware that simply ignores the PPS signal. This realization cost me half an afternoon – it seemed to drag on endlessly if you ask me. On the other hand, time flies – but not always, apparently.