On Mon, Oct 7, 2013 at 6:32 AM, Marco Tedaldi <[email protected]> wrote: > Now, since I have to set up ntp on all the guests anyway, I've been thinking > about > how to best configure the system. I've found out, that the systems can be > configured > as peers. > > What's the difference if I configure the servers as peer or if I just put all > the servers > into each other configuration files as servers?
I do not think you would want virtual machines running on the *same* host to peer with one another. They share too much physical and virtual hardware for that to be a reliable configuration. Especially when you consider that NTP peers following the same hardware timer on the physcial host would "outvote" an off-host NTP source that is actually more correct. You would not want mutual "server" configuration lines for the same reason. The potential failure modes are ugly. Ideally, I think you would want all the VMs to be clients of some set of off-host NTP servers. We run hundreds of VMs this way on VMware ESX hosts, and have had no issues with VM timing in many years. Even if those off-host NTP servers are across a WAN connection (such as pool servers), it will still likely be more reliable than using virtual machines as a time source. You really want your time sources to be "real" physical systems as much as possible, as virtualization inherently offers much worse timing than an OS running on bare metal. This is because the VMs themselves are time-sliced across the real CPUs, and so they can have clock ticks and other interrupts delayed by the hypervisor, leading to more jitter and even long-term drift. -- RPM _______________________________________________ pool mailing list [email protected] http://lists.ntp.org/listinfo/pool
