Configure ntpd to point to the usual sources, but also peer with the ntpd you are running in other regions. Enable statistics. You now can record the offset of instances in different regions.
Do not expect sub 10 ms error or so between regions. The latency between regions, or to distant public NTP servers, makes that difficult.
Your requirements are probably not as extreme as one of the most distributed databases, Google Spanner. They developed a time service just for inter-datacenter databases, added hardware atomic clocks, and still I think ε was in the single digit ms.