The kernel clock updates the CMOS clock from kernel time every 11
minutes, and it does this within the time window of one tick.
Unfortunately the update of the CMOS clock's time also affects the
periodic interrupts. I've seen jumps up to 80ms. This seems to be
interesting, especially as on would expect only +-5ms (10ms is one
tick on i386).
This in turn affects everyone who uses /dev/rtc to sample some data
at a high rate (>128 Hz).
For my problems with NTP, I've found out that a tick value of "10001"
instead of "10000" will make things better, but this might depend on
crystal and prescaler of the motherboard. Tests are still running...
Ulrich