Re: [Ntp] Antw: [EXT] Minutes from IETF 108 NTP WG session

[Miroslav Lichvar <[e-mail]>]

On Mon, Aug 03, 2020 at 08:17:08AM +0200, Ulrich Windl wrote:
> "It turns out that for trades a nanosecond resolution matters."
> 
> I wonder what's really behind that claim: Reading the current system time on a
> current Linux server yields in about 1µs precision (independent from how the
> clock is synchronized).

On modern computers running Linux, at least for the last decade, the
time it takes to read the system clock is usually less than 50
nanoseconds. I have not seen a sub-10ns clock yet. That's still an
order of magnitude to get into the sub-nanosecond range. Even if the
CPU was so fast that it could read and convert the TSC counter in less
than a nanosecond, there is no POSIX API to return the timestamp in a
sub-nanosecond resolution.

I guess for hardware clocks (e.g. NICs that support hardware
timestamping) a sub-nanosecond resolution might be needed sooner than
for the system clock.

NTP has a 1/4ns resolution. For the White Rabbit that might not be
good enough. For synchronizing ordinary computers (in financial
industry or other) I think that will be sufficient for quite some
time. In any case, a correction extension field (which will need to be
supported in switches/routers in order to get a sub-nanosecond
accuracy) can have a better resolution than the timestamps in the NTP
header.

> So before claiming that financial industry needs sub-nanosecond resolution,
> I'd like to see hard facts and details.

FWIW, I have not heard anyone asking for a sub-nanosecond resolution
yet. The accuracy of the clock is a bigger issue.

> The other thing is: How small would the stability of the clock have to be to
> make any sub-nanosecond resolution be realistic? 0.001 PPM?
> Real "normal" servers will probably have 1 PPM, and a "certified chronograph"
> will allow around 50 PPM (AFAIK)...

For sub-nanosecond stability with an ordinary computer clock, the
update interval would need to very short, e.g. few milliseconds.

In a test I did with two directly connected idle computers and a
polling interval of -6 (using the interleaved mode), the reported
stability was just a few nanoseconds. With a shorter interval it might
be possible to get it under 1 nanosecond, but I'm not sure what would
be the point when the precision and accuracy of the clock is in tens
nanoseconds at best.

-- 
Miroslav Lichvar