Linux Virtualization - Jun 2016 - Heavy clock drifts inside of KVM guests (qemu-kvm 2.1.2)

Dear list(s),

we are running a cluster of virtual machines on ganeti 2.15.2 on top of
qemu-kvm 2.1.2. Hosts are Debian jessie (Kernel 3.16). We have big trouble
with timekeeping on guests, no matter if these guests are lenny, squeeze,
wheezy or jessie. Clock drifts heavily on some, others are fine. There is
no obvious pattern on which guest this happens and on which not, but it
seems the clock drifts heavier on guests that are idle and have nothing to
do. Some more facts:

* Clocksource on all hosts are set to tsc
* Clocksource on all guests are set to kvm-clock
* CPUs on the hosts are Intel Xeons, all have these related flags: tsc,
constant_tsc, nonstop_tsc. One host has tsc_deadline_timer as an extra,
because it's a newer Xeon, but it doesn't seem to make a difference in
drifting.
* Most guests complain about "Clocksource tsc unstable" in dmesg after
boot
(even when set to clocksource=kvm-clock)
* The hosts don't drift at all, we only see the issue in guests
* We don't use CPU pinning for KVM guests
* We run OpenNTPD 5.7p4 on both the hosts and the guests (we build this for
ourselves, so the version is the same in every guest across every Debian
version)

We tried different things to debug the cause of this issue but we're out if
ideas now. This is what we have tested so far:

* Newer kernel from jessie-backports
* Newer qemu-kvm from jessie-backports
* Set clocksource=hpet on the hosts, this made drifts even worse (not
tested yet: acpi_pm)
* Tried to run guests without kvm_steal_time setting
* Tried without running OpenNTPD on the guests

We have some hints along our investigations. Novell for example says that
CPU pinning may help on drifting guests:
https://www.novell.com/support/kb/doc.php?id=7008698

Others suggested using an "failover clocksource" on the guests:
https://blog.laimbock.com/tag/clocksource_failover/

However, before we just continue to poke around in the dark, we wanted to
hear input from virtualization experts that may have stumbled across
something like this before. Any input would be greatly appreciated!

Please reply directly, as I'm not subscribed to the list. And sorry for
cross-posting this to several lists at once, we just hope to get as much
feedback as possible.

Thanks in advance,
Frank
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20160621/2f03d246/attachment.html>