Re: [Linuxptp-devel] [PATCH RFC 0/6] Improve accuracy with software timestamping

[Keller, J. E <jac...@in...>]

On Fri, 2015-02-13 at 13:56 +0100, Miroslav Lichvar wrote:
> These patches should improve the synchronization of the clock with
> larger jitters, e.g. with software timestamping, wireless networks,
> etc.
> 
> The idea is to give smaller weights to samples where the sync and/or
> delay messages were delayed significantly in the network and possibly
> include a large error. The sample offset is based directly on the
> sample delay instead of the long-term filtered delay and the sample
> weight is set to the ratio of the sample delay to the long-term
> average. E.g. if the measured delay is normally 1 ms and the new
> sample has 10ms delay, the sample weight (and the clock adjustment
> that will be made) will be 10 times smaller.
> 

This seems like a good idea, to help in the case where spiked delays
could cause swings which take a while to settle out. A bit like dropping
the outlying parameters.

> This "weighting" mode is enabled only when the sync interval is equal
> or longer than the delay/pdelay request interval to not reuse the
> delay timestamps too many times and keep the interval between sample
> times stable. In some tests I saw an improvement also with 2:1
> sync/delay rate and maybe even higher, but it would probably be tricky
> to implement if it should be enabled automatically, possibly requiring
> some cooperation with the delay filter.
> 
> In my testing I did so far it seems to work nicely. With the linreg
> servo there should be no regression. With PI there can be worse
> performance observed when the constants are not configured properly,
> that is when the servo is too slow to track the frequency changes and
> weighted samples make it even slower. Currently, the weighting mode is
> enabled only when ki <= 0.01 to include the default SW timestamping
> constant and not the HW constant. It could be always enabled or I
> could add an option to override it if you think it would be useful,
> I'm not sure.
> 
> What do you think? Does this make sense?
> 

I like the idea, but I confess you do know more about the clock than I
do.

Regards,
Jake

> 
> Miroslav Lichvar (6):
>   clock: save delay timestamps and correction.
>   clock: split out calculation of sample delay.
>   servo: add support for weighted samples.
>   clock: set sample weight.
>   linreg: use sample weight.
>   pi: use sample weight.
> 
>  clock.c         | 85 ++++++++++++++++++++++++++++++++++++++++++++-------------
>  clock.h         | 10 +++++--
>  linreg.c        | 64 +++++++++++++++++++++++++++++--------------
>  ntpshm.c        |  1 +
>  phc2sys.c       |  2 +-
>  pi.c            | 20 ++++++++++++--
>  port.c          | 10 +++++--
>  servo.c         | 11 +++++++-
>  servo.h         | 10 +++++++
>  servo_private.h |  4 ++-
>  10 files changed, 169 insertions(+), 48 deletions(-)
>