[ntp:questions] Running NTP servers with different time offsets relative to some common "root" server?

[Tobias Gierke]

Hi everyone,

I'm facing a NTP-relateded challenge at work and hope that somebody on 
this list has maybe solved a similar problem already:

- Our product is a system consisting of multiple components that may or 
may not run on the same host and all parts require time synchronization 
at all times because we're processing timestamped measurement data
- Our application does a lot of heavy computation triggered at the start 
of every minute (by crond)
- Since a lot of different application versions are deployed in the 
field and customers are slow to upgrade, we need to test many different 
versions of our application in our lab
- To ease hardware requirements for testing we run all those different 
versions of our application inside virtual machines, currently all 
synced to the same NTP server
- Since all those VMs are synchronized to the same server, all cron jobs 
on those VMs kick in at the same time, overloading the VM host periodically

I already looked into a simple solution (like being able to pass some 
magic "--offset XX seconds" option to ntpd) but it seems there is no 
such thing.

So my current idea is to

1.) have 4 different time servers that each are offset from all the 
others by 15 seconds (so server #1 would have offset 00 seconds, server 
#2 would have offset 15 seconds, etc.)
2.) configure those servers with ULC w/ PPS and hook-up the same PPS to 
all 4 servers OR have one PPS source for each server and have all of 
them sync to a common source (like a DCF77 time signal)

My problems are with the second step: I don't have a PPS source, I only 
have a Meinberg DCF600 USB 
(https://www.meinbergglobal.com/english/products/usb-dcf77-clock.htm), 
but I couldn't find any documentation on how to use it as  a PPS source 
only (and ignore the actual "time" part of the synchronization). 
Furthermore I don't really want to buy 3 more of those clocks if I can 
help it (especially considering the fact that we might need even more 
different "time domains" as the number of VMs increases).

Any ideas?

Thanks in advance,
Tobias