NTP Reference Clock Statistics
Update -- Autumn 2009
Unfortunately, my stable of NTP servers is in some disrepair right now, and due to a likely relocation in my future, probably won't be running well for the forseeable future. So, the statistics here probably aren't very useful, but hopefully the information will still be of interest.
Note to WIRED readers:
If you found your way here as a result of the article about "time-nuts" at wired.com, I need to clarify the quotes about my NTP servers keeping time to 100 nanoseconds. While my oscillators (Cesium, Rubidium, and disciplined) normally track each other within about that range, you can't really get nanosecond accuracy over a 100 Mbit ethernet, so the performance of the NTP servers, when seen across the wire, doesn't match that of their reference sources. As you can see below, when things are well behaved the normal noise across the ethernet is 10 to 20 microseconds, with occasional spikes that are greater; I'm fairly certain those are due to network delays.Also, to make measurement easier a number of the clocks, including the Cesium ones, are deliberately offset from "correct" time by a few tens of microseconds.
Finally, the WIRED article says I make my time service to anyone who wants it. That's true after a fashion, but I don't have the bandwidth to run a wide-open server. If you're interested in precise timing and would like to sync one of your machines to my clocks, please drop me an email and I'll get you set up.
I run four NTP Stratum 1 servers with attached reference clocks. Below this introduction are graphs that show performance statistics about them, as well as some other internal and external NTP servers that I monitor.
tick, tock, toe, and cesium are dedicated NTP servers, built from Soekris net4501 single-board computers with 133MHz AMD "Elan" processors (486 class). That may not seem like much horsepower, but these little guys have a secret weapon: a very high resolution timer that can be used to get timekeeping accuracy far better than any normal PC.
In addition, these Soekrises (Soekri?) have been modified to replace the normal crystal oscillator with a synthesizer driven by an external high stability source which makes them almost immune to temperature variations and other errors.
I've written up details on how to convert the net4501 computers for high performance timekeeping.
Here are details about the four servers:
- cesium gets its PPS and its clock from an HP 5061A Cesium Beam
Frequency Standard (a real "atomic clock"). Because it does not rely on a
radio receiver, cesium has low jitter and is not dependent on
external factors. So, I use it as the reference system that monitors the
other servers. (Refid "ATOM")
- tick syncs to the PPS from an Austron 2100T LORAN-C receiver
which is synchronized to UTC. Its clock comes from an Austron 2010B
disciplined oscillator which is steered by the 2100T. (Refid "LOR")
- tock is connected to an HP Z3801A GPS-disciplined oscillator.
It syncs to both the serial data stream and the PPS output, and its clock
source is the disciplined oscillator in the Z3801A. (Refid "GPSZ")
- toe is connected to a Spectracom 8170 WWVB receiver. It syncs
to both the serial data stream and the PPS output. Its clock comes from a
companion Spectracom 8164 WWVB disciplined oscillator. (Refid "WWVB")
While toe has the advantage of being directly traceable to NIST time through WWVB, it is actually the least stable of the clocks in the short and medium term because ionospheric conditions affect the WWVB signals as they propagate from Boulder, Colorado to my house. It often wanders by a few milliseconds, and when the signal momentarily disappears completely, it can spike 20 or 30 milliseconds before recovering.
The charts below show the offset of each of the internal servers, as well as some internal and external servers, referenced to cesium. All the internal machines are on a 100 Mb Ethernet LAN, with a couple of switching hubs in line. There's also a chart, for what it's worth, showing what cesium thinks is the offset of the HP 5061A clock that drives it.
The "current day" charts show performance for the current day since 0 UTC (updated hourly). The next set of charts shows the preceding seven full days, plus the current day. The final charts show 30 days of data. (Here are the tools I used to generate these plots.)
One of my externally-visible stratum 2 servers, meow.febo.com is part of the pool.ntp.org project. Statistics about meow are here.
Current Day
Last Week
Long Term
