Understanding clock drift in virtual machines w/ CFS, NOHZ

Understanding clock drift in virtual machines w/ CFS, NOHZ

[Mika Bostrom]

  Good day.

  I have come across the occasional issue of clock drift when running
(earlier) kernels inside a virtual machine. The reason for drift is
simple enough - host may be under heavy load, guest may not get enough
resources to run all the ticks (HZ) it is supposed to. At least to put
it simply.

  Now, I have been wondering how the introduction of CFS and CONFIG_NOHZ
(dynticks) change this, or if they do at all. If I have understood
correctly, only the host's scheduler is really involved so it enters the
picture when host is running a recent kernel. In a pure hypervisor-only
virtualization, that point should be moot.

  And then we have dynticks. Try as I might, I haven't been able to wrap
my head around the combination. What happens to the kernel and
timekeeping when guest has NOHZ enabled?

  My google-fu is not good enough to find the relevant documentation if
such even exists on this subject. To top it off, I'm not certain whether
my question is trivial, non-issue or just plain weird.

  Any pointers on where I should look for more information will be
appreciated.

-- 
 Mika Boström      +358-40-525-7347  -=-  The flogging will continue
 Bostik@iki.fi    www.iki.fi/bostik  -=-  until morale improves
 GPG: 0x039F188E; EC67 5B3A E6E3 6A84 9CB2  94D3 BFCD BD57 039F 188E