I said:
I also wonder whether it might be productive to consider closing the NTP "servo loop" in velocity (frequency) in this case, rather than position (phase).
Before somebody else calls me on it, I should point out that NTP actually uses both: "The clock discipline algorithm functions as a nonlinear, hybrid phase/frequency-lock (NHPFL) feedback loop." (see http://www.cis.udel.edu/~mills/database/brief/clock) I'm wondering whether the clock discipline might better emphasize only frequency in such circumstances, as opposed to merely tempering the PLL. Rob