Re: Improving accuracy of PHC readings

[Miroslav Lichvar <mlichvar@redhat.com>]

On Mon, Oct 22, 2018 at 03:48:02PM -0700, Richard Cochran wrote:
> On Fri, Oct 19, 2018 at 11:51:37AM +0200, Miroslav Lichvar wrote:
> > The extra timestamp doesn't fit the API of the PTP_SYS_OFFSET ioctl,
> > so it would need to shift the timestamp it returns by the missing
> > intervals (assuming the frequency offset between the PHC and system
> > clock is small), or a new ioctl could be introduced that would return
> > all timestamps in an array looking like this:
> > 
> > 	[sys, phc, sys, sys, phc, sys, ...]
> 
> How about a new ioctl with number of trials as input and single offset
> as output?

The difference between the system timestamps is important as it gives
an upper bound on the error in the offset, so I think the output
should be at least a pair of offset and delay.

The question is from which triplet should be the offset and delay
calculated. The one with the minimum delay is a good choice, but it's
not the only option. For instance, an average or median from all
triplets that have delay smaller than the minimum + 30 nanoseconds may
give a more stable offset.

This is not that different from an NTP client filtering measurements
made over network. I'm not sure if we should try to solve it in the
kernel or drivers. My preference would be to give the user space all
the data and process it there.

If the increased size of the array is an issue, we can reduce the
maximum number of readings.

Does that make sense?

-- 
Miroslav Lichvar