[mkl]

The problem is using watch you have no control* over when in each second it's getting the time, so it could be nearly a second late (e.g. it's getting the time once per second, but happens to be doing it when it's a few milliseconds away from ticking over to the next second).

* Okay, you have a little control in that you can press enter, or otherwise set it running, at a particular moment.

[Izkata]

So don't have any delay:

  while true; do date +%s%N; done

[mkl]

There's still a delay with that, as date takes time to run. You need a program that gets the current time, waits an appropriate amount, prints, and repeats. It'll still be slightly off unless you can measure the printing and flow control accurately.

I have a program I use in shell scripting called sleepuntil that does something like this, but it doesn't try to be millisecond-accurate.

[nonrandomstring]

The problem that I noticed is if you open two terms on different hosts (say by ssh) and run watch commands then, regardless of resolution, they'll be out of sync. There's no hard tick event common to both kernels. And as someone already said, you introduce sampling effects.

So I figured you need a time "diff", a single command that updates from both timeservers and then presents the offset in a single operation. So (I just researched this again and) there's three answers here [0], using ntpdate and something I've never seen before called clockdiff.

[0] https://stackoverflow.com/questions/2296981/how-can-i-work-o...

[mkl]

Right, as I said, watch is the wrong tool here. The event in common is the synchronised clocks, or an NTP server.

To see the NTP offsets of a machine you can run something like:

  ntpq -pn

ntpdate -q doesn't seem very consistent to me, even pointing it at a nearby server.

To see the clocks ticking visually you can do something like the following and run it on each machine (assuming they have the same ping).

  #include <time.h>
  #include <sys/time.h>
  #include <stdio.h>
  
  void main() {
      struct timeval tv;
      struct timespec ts;
      ts.tv_sec = 0;
      while(1) {
          gettimeofday(&tv, NULL);
          ts.tv_nsec = 1000000000 - 1000*tv.tv_usec; //number of nanoseconds left in the current second
          nanosleep(&ts, &ts);
          gettimeofday(&tv, NULL);
          printf("%lu.%06lu\n", (unsigned long) tv.tv_sec, (unsigned long) tv.tv_usec);
      }
  }

Note that the sleep and second gettimeofday call take a little time (between 70 and 300 μs for me, but it depends what else is running), so the tick times reported won't be exactly on the second.

[nonrandomstring]

Great detailed answer. I upvoted you.