[LinuxPPS] Linux 2.6.28 - PPS jumps

[Heiko Gerstung]

Hi Guys!


Thanks for all your replies, I am still working on this and my current 
status is that I see sudden jumps - now even with the PPS reference 
clock - of around 30 microseconds.

The loopstats records during such an incident look like this:
54853 21913.755 0.000001352 23.933 0.000001907 0.000129 2
54853 21916.755 0.000001442 23.933 0.000001907 0.000123 2
54853 21919.755 0.000001601 23.933 0.000001907 0.000117 2
54853 21924.756 0.000001491 23.933 0.000001907 0.000115 2
54853 21928.752 -0.000290294 23.915 0.000103177 0.006265 2
54853 21934.753 0.000004655 23.915 0.000142088 0.005861 2
54853 21937.753 0.000029043 23.917 0.000133191 0.005503 2
54853 21941.754 0.000028558 23.918 0.000124591 0.005184 2
54853 21947.754 0.000026432 23.920 0.000116546 0.004883 2
54853 21953.755 0.000023054 23.921 0.000109026 0.004595 2
54853 21958.755 0.000022256 23.923 0.000101986 0.004325 2
54853 21963.756 0.000021369 23.924 0.000095402 0.004071 2

As you can see, I patched ntpd to allow for a higher polling rate and 
during this experiment, the PPS reference has been polled every 4 seconds.

This hickup can be observed 3 times in the last 7.5 hours and it looks 
like this happens in a fixed interval:
54853 2363.752 -0.000303320 [first incident this morning]
54853 12146.752 -0.000345740 [after 9783 seconds]
54853 21928.752 -0.000290294 [after 9782 seconds]

This looks *extremely* equally distributed to me. No need to say that 
the PPS of this device worked fine with 2.6.15.1 and ntp-4.2.0b ... I 
have no idea where this might come from and will continue testing ... 
IMHO the 163 minutes between two incidents are not a strong indication 
for a kernel issue. I also do not think that the polling rate will 
affect the chance that this happens, if this would be a question of 
chance, I can not imagine that I am "lucky" every 9780 seconds.

The next incident - if the 9870s interval is maintained - is "scheduled" 
in ~1 hour and I will try to check how the value in the 
/sys/class/pps/pps0/assert file looks like during such an event. But due 
to the NTP filtering algorithms I suspect that this is more a problem of 
the system time or frequency jumping, because the time jumps 29 
microseconds into one direction (-29us), NTP corrects for that (next 
value is +4 us) and then the time jumps back (next offset is +29us), 
causing NTP to slowly correct the time.

Any ideas where to look? Did anyone ever spot something like this?

Regards,
Heiko