Re: [gpsd-dev] [PATCH 1/1] Initial draft; how to estimate time1 offset

[Andy Walls]

On Thu, 2013-11-07 at 10:04 +0800, Sanjeev Gupta wrote:
> 
> On Wed, Nov 6, 2013 at 8:20 PM, Greg Troxel <address@hidden> wrote:
>         My impression is that fudge1 is added to the offset, and that
>         for
>         typical timecodes (e.g. ublox binary), a fudge of around
>         0.110s is
>         normal, indicating that the timecode arrives 110 ms late,
>         which would
>         produce an offset of -0.110s.
>         
>         It would be nice to go through this carefully with the signs,
>         and to
>         usea value that appears in practice.
>         
>         It would also be nice to collect typical values.
> 
> 
> Greg, Andy,
> 
> Please see the enclosed image.  The righthand side is before midnight,
> the left is today (I merged the peerstats).
> 
> 
> Assuming I have decided that the value is -0.47 , what will be my
> time1 value?  -0.47 or +0.47 ?


+0.47.

IIRC, ntpd references everything to the system clock.  So here is the
relationship

        offset_clockN = time_clockN + fudge_time1_clockN - time_system_clock


For now assume the time_system_clock is "truth".

The objective is to have the average offset_clock0 == 0, so we take the
observed average offset from the peerstats for clock0 and combine it
with the fudge_time1_clock0.

        0 = time_clock0 + (fudge_time1_clock0 - offset_clock0) - 
time_system_clock
        0 = time_clock0 + fudge_time1_clock0' - time_system_clock

The new

        fudge_time1_clock0' = fudge_time1_clock0 - offset_clock0

If you started your calibration test with

        fudge_time1_clock0 == 0

then the new

        fudge_time1_clock0' = -offset_clock0


------

But we know that the system_clock really isn't truth, because we know
the PPS is probably better.  For a better fudge_time1_clock0, we want to
reference to the PPS (clock1) and not the system_clock: 

        offset_clock0 = time_clock0 + fudge_time1_clock0 - time_system_clock
        offset_clock1 = time_clock1 + fudge_time1_clock1 - time_system_clock

Subtracting those two yields:

        offset_clock0 - offset_clock1 = 
                        time_clock0 + fudge_time1_clock0 - (time_clock1 + 
fudge_time1_clock1)

We will assume that fudge_time1_clock1 (for the PPS) is a constant that
is correct.

We have an objective to make the average offset of clock0 relative to
clock1 == 0, so

        0 = time_clock0 + (fudge_time1_clock0 - offset_clock0 + offset_clock1)
                                - (time_clock1 + fudge_time1_clock1)

        0 = time_clock0 + fudge_time1_clock0' - (time_clock1 + 
fudge_time1_clock1)

The new

        fudge_time1_clock0' = fudge_time1_clock0 - offset_clock0 + offset_clock1

If you started the calibration with 

        fudge_time1_clock0 = 0

then the new

        fudge_time1_clock0' = -offset_clock0 + offset_clock1


Again, the offset_clockN numbers are stable averages you extract from
the peerstats file for clock0 and clock1 respectively, and that clock1
is the PPS which we trust as truth.

        
Regards,
Andy