[clug] October Canberra Linux Users Group meeting
bob at cs.anu.edu.au
Wed Oct 21 19:07:52 MDT 2009
Michael James wrote:
> On 21/10/2009, at 3:40 PM, Alex Satrapa wrote:
>> Please bring your laptops and experience/curiosity of virtualisation!
> The biggest problem I've had with virtualisation is timekeeping.
> On real hardware it's been a solved problem since the 90s.
> Run ntpd with a good many servers in a loose network.
> Kept the clocks accurate and was robust, using consensus
> meant any server that went haywire was ignored.
> In virtual bottles having the clock counting ticks doesn't work,
> as the machine isn't there all the time and misses ticks.
> Opensuse and SLES are bad at this, as they run the clock at 1000Hz.
> In a guest, ntpd's assumptions about drift don't hold.
> Running ntpd in multiple guests is an even bigger no-no
> as they seem to fight over the clock.
> ?Does setting the hardware clock in a guest touch the hosts clock?
> Anyway I've tried to run ntpd once in the underlying host OS,
> but I haven't found the way to feed that time up into the guests
> as an absolute value, not ticks that can be mis-counted causing drift.
> My 2 seconds worth,
I've been playing with OpenVZ a bit recently (www.openvz.org) - which is
OS level virtualization. It's really the Linux version of Solaris's
containers and possibly a little more featureful than BSDs jails.
In OpenVZ there is only one kernel and if it can keep time (with the aid
of, eg., ntp) then all the "guest"s (actually "Virtual Environments" -
VEs) get the correct time for free.
Interesting thing is that the OpenVZ documentation recommends against
running any services other than SSH in the "Hardware/Host Node" (HN)
and instead running ntpd in one of the VEs. This VE then needs special
permission to be able to set the system time on behalf of the HN and
the other VEs. Seems to work well enough.
More information about the linux