There doesn't seem to be anything wrong with your NTP config or with the configured peer, so maybe the local clock in one or more of the cells is occasionally drifting too far too quickly and going beyond the range where NTP can correct it.
Adding the following to (the top of) your cells ntp.conf might be of help:
tinker panic 0
This setting allows ntpd to accept any offset and is considered a VMware timekeeping best practice for linux guests -- http://kb.vmware.com/kb/1006427
More extreme measures might be to use a different local clocksource or to reduce the maxpoll frequency to force ntp syncs to occur more often (adjusting maxpoll is usually not required/recommended).
Cheers
Mark