Agreement isn’t accuracy

I finally got my GPS-locked NTP server on the network. For grins, I added a few of my servers to its `ntp.conf` file for monitoring.

-> ntpq -p
     remote           refid      st t when poll reach   delay   offset  jitter
==============================================================================
*GPS_PRAECIS(0)  .GPS.            0 l    5   16  377    0.000    0.000   0.002
+140.99.199.146  17.253.2.123     2 u   58   64  377   23.931    3.811   1.232
 SyncServer.loca 17.253.2.123     2 u   62   64  377    3.240    3.658   0.726
+nyc.ft.primeser 17.253.2.251     2 u   10   64  377   21.271    2.328   3.119

At a glance, it seems like there’s pretty good agreement than an offset of around +3ms is needed here. Naively, I might glance at this and assumed the middle two were best, and think that 3.658-3.811ms was probably right. (The SyncServer has the noselect keyword applied.)

But the local clock is locked to GPS, and the 0.000 offset is correct. The “agreement” on a +2-4ms offset is wrong, and probably reflects some routing asymmetry with Comcast. (Or it’s just random coincidence.)

(It’s amusing to me that three different NTP servers of mine in different parts of the Northeast have all selected Apple’s NYC NTP servers as sources. It’s really just coincidence, though they are well-connected.)