Datacenter GNSS Time Best Practices
Eight things worth doing, from the closing slide of a talk on getting the last nanoseconds to UTC. Most cost nothing but attention, and most of the error they remove is invisible until somebody checks.
Tagged “timing”
31 pages · all tags
Topics In Timekeeping
What are the limits of GNSS time accuracy?
The limits are not about radio or geometry. GNSS time is a prediction, so its accuracy is bounded by how good the corrections are — and corrections are forecasts of errors that have not happened yet.
What makes an accurate GNSS timing receiver?
Six traits, in the order they start to matter. The last one separates a box that wants to be your master clock from a box that knows you already have a better one.
Choosing a GNSS timescale for syncing datacenter clocks
You are not choosing a receiver so much as choosing whose timescale you end up on. There are three ways to do it, they differ by a factor of a hundred in cost, and most datacenters should take the cheapest.
How does GNSS holdover work?
When GNSS goes away your clock coasts on the frequency it was last told. How far it drifts is decided by your room more than your oscillator — which is why a holdover measurement and a holdover specification are different objects, and only one of them is a promise.
Can I build my own link to UTC(NIST)?
The data you would need is public, free, and better than most people realize. You can build the measurement in a weekend. What you cannot build is the one thing that turns it into a number somebody else has to accept.
Is my clock right?
Timing failures are usually quiet, and nothing inside a wrong clock knows it is wrong. The only detector is a second clock — and the boundary between what the two share and where they are independent decides what you can learn by comparing them.
Comparing distant clocks
You cannot run a cable, and neither clock can be the source for the other. So you both watch a third clock in the sky — which drags its error into your answer. Every technique here is a different way of removing it.
Questions to ask a clock vendor
Nine questions whose answers tell you whether a vendor understands their own product — no test equipment required.
Time Distribution Protocols
NTP, NTS, PTP, White Rabbit and a bare PPS, compared on the two things that decide which you need: how close they can get, and whether they account for the length of the wire.
Questions to ask when benchmarking clocks
You have the hardware and you want to compare it. Most of the traps are in the comparison, not the clocks.
Ways to get time into a datacenter
GNSS, a fiber from a national lab, a commercial delivery service, or NTP from the internet. Why GNSS usually wins below 100 ns, and why you want more than one.
Can I sync my datacenter clocks to UTC?
Literally no, but practically yes. UTC is computed weeks after the fact, so nothing on Earth can synchronize to it — and yet a well-built datacenter clock sits a few nanoseconds away. Here is how both of those are true.
What are timestamps and timescales?
A timestamp is a count, and a count means nothing until you say what you counted and from when. That is a timescale: tick length, origin, and count — and anything with all three qualifies, including some absurd ones.
How big is a nanosecond?
A nanosecond is about a foot. Once the units are physical rather than numerical, most timing arguments get easier to have.
How does GNSS time relate to UTC?
Not one relationship but two: a whole number of seconds you can look up, and a sub-nanosecond prediction that two control loops — one running in minutes, one in weeks — spend their lives maintaining.
Matching acquisition and distribution
Two distinct problems that both have to be solved. Their errors sum, so the accurate one is paying for accuracy the other cannot deliver.
Which GNSS constellation keeps the best time?
Folklore says Galileo. BIPM measures it daily, per constellation, and publishes the answer — and over fourteen months the folklore does not hold up.
How big is a degree?
Each decimal place of latitude and longitude buys a factor of ten. Knowing where the useful digits stop saves arguing about the ones that cannot mean anything.
UTC(k), and who defines yesterday
The realizations you can actually reach in real time — and why the laboratories with the best clocks get the most say in what UTC turns out to have been.
What does "traceable to UTC" actually mean?
It means an unbroken chain of comparisons back to UTC, each with its uncertainty stated. Not a logo, not a certificate, and not a GPS antenna — which is why most systems that claim it do not have it.
Why do you want precise time?
Precise time is seldom the goal. It is the price of admission for something else — which is why the requirement usually arrives from outside engineering.
How to configure antenna position for datacenter GNSS receivers?
If you don't know your exact location, you can't know the exact time — about a nanosecond of error for every foot you are off vertically. Here is how to get it right.
Accurately measuring elapsed time
A clock so bad the internet refuses to talk to it still measures a microsecond-scale interval to about a nanosecond. Short durations forgive both ways a clock can be wrong, and knowing when you are measuring one saves real money.
How to configure feedline length for datacenter GNSS receivers?
An uncalibrated feedline is likely your largest source of error to UTC — tens of nanoseconds from a cable nobody measured. Here is how to find the number and configure it.
Visualizing precision, trueness, and accuracy
The most-drawn diagram in metrology, with the axis everyone mislabels put right. Trueness and precision are the two axes; accuracy is the corner where you have both, not a third thing.
Measurement resolution and accuracy
Coarse resolution does not merely limit precision — it counterfeits it. A grid too coarse to show scatter returns the same confident number every time, and stops being able to warn you.
The limits of averaging
Averaging buys you improvement that stops. What is left when it stops is bias — and the same story plays out in position, in time, and in frequency.
Do timestamp errors average out?
No. The archer gets ten arrows and can average them; a network packet is timestamped once. That difference decides which errors you can tolerate — and it moves all the work to before the event.
Should I prioritize clock agreement or accuracy to UTC?
Usually a false choice. The two are not opposed, most improvements buy both, and the real tradeoff only appears at the very last nanoseconds — where it is a genuine engineering preference rather than a confession of low standards.
From the blog
Topics In Timekeeping Is Live
Thirty pages on precise time in datacenters — what limits GNSS accuracy, what traceability actually requires, and how you would know your clock was still right. Not blog posts. Pages I intend to keep correct.