[ntp:questions] R: Re: debugging strange ntp in virtual environment

David Woolley david at ex.djwhome.demon.invalid
Thu Sep 19 20:53:34 UTC 2013

> What means simulated real time can run at "variable rates"
> ?!

This is explained in one of the early VMWare white papers on timing (at 
the time they were recommending using VMWare Tools to control the guest 
time and synchronising the host.  There seem to have been changes since, 
but I haven't seen a rationale for the change, which may well be just 
the best of a bad job.

Although in the long term the host will simulate the correct number of 
timer interrupts to the guest, it may actually miss several, and then 
send a burst, to minimise the cost of switching to the guest.  That 
means that simulated time will run very slowly, and then run very fast 
over a few milliseconds.

As ntpd is trying to time stamp events that occur on a smooth timescale, 
it is likely to get confused.
>> Both can result in very large jitters and make it difficult for ntpd to
>> work out the underlying clock frequency error.
> Ok, so you suggest to run ntpd
> service (to give time to clients) on physical machine ?

I suggest that, for any application where you need to prove 
non-causality, or anything else with tight timing constraints, you need 
to use physical machines for all elements of the timing chain, both 
servers and clients.  As servers may get shared by systems with 
different requirement, they generally have to meet the highest requirement.

By "proving non-causality", I mean the typical financial dealing 
requirement to show that a trade could not have been influenced by 
another trade.

More information about the questions mailing list