Richard,

We are using ptp to train an oscillator for a satcom applicationn. We
have a requirement that the oscillator frequency be correct to a
tolerance of 5 PPB.

Our board has a Marvell PHY with hardware time stamping. The time
stamping counter is driven by the oscillator that is being trained.
The oscillator (AD9545) is pretty stable and is adjustable to a small
fraction of one billionth.

When we slave our board to a GPS, we see from the log that the offset
is very stable, and very close to zero for long periods of time.

But in real life, our board will not be directly connected to a GPS.
Our board will be slaved to a boundary clock in a Mellanox 2410
switch. When our board is slaved to the Mellanox, we see considerable
variation in the offset reported in the log. The offset variations are
big enough that it causes us to go outside our 5 ppb frequency budget.

Mellanox manuals imply that the switch uses hardware time stamping,
but I suspect that is not completely true, or that the granularity is
not close to one ns.

As an experiment, we patched pi.c to do exponential smoothing to the
offset before using it in the new frequency calculation. With a
smoothing factor of 0.0625, this reduced the frequency error to less
then 2 PPB, well within our tolerance.

It occurs to me that our use case is not contemplated in the
design of PTP. Normally, ptp syncs a lower-quality clock to a
higher-quality one. Our case is the opposite.

I understand that if the t1 time stamp is accurate, the delay filter
should provide all the smoothing you would need. I just question the
accuracy of that timestamp (though I think the error is just jitter
that gets cancelled out by subsequent errors in the other direction.
Since the Mellanox is slaved to the GPS, there is no long-term
accumulationof error).

Do you have any comments or suggestions?

Pat O'Neil
Hughes Network Systems
Germanttown, Md. 20876
on...@hughes.com
_______________________________________________
Linuxptp-devel mailing list
Linuxptp-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linuxptp-devel

Reply via email to