On Fri, Aug 29, 2014 at 06:29:54PM +0200, Kay Sievers wrote: > On Wed, Aug 27, 2014 at 4:47 PM, Miroslav Lichvar <mlich...@redhat.com> wrote: > > The shortest time constant that is stable with the kernel PLL (compiled > > with SHIFT_PLL=2) is about log2 of update interval - 3. Set the constant > > to poll - 2 to make room for one missed update. > > Experimentation did show, that the adjustment of the kernel was too > gentle to reach the target time in the desired window, so we made it > more stiff.
That's the problem with PLL, it adapts to frequency changes very slowly. > All timesyncd really cares is that we are not jumping, we > do not really need to smooth out the adjustments. It would probably help if the frequency was estimated over a fixed interval on start before enabling the PLL mode (similarly to ntpd), or at least the frequency was restored from a drift file, so it doesn't always have to start from zero. > We would re-check what happens with your patch, I'll hold back > applying this for now. It should take about four times longer to converge to zero offset. -- Miroslav Lichvar _______________________________________________ systemd-devel mailing list systemd-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/systemd-devel
