On Thu, Sep 06, 2012 at 07:44:13AM -0400, Laszlo Hornyak wrote:
hi!
Related to this patch:
http://gerrit.ovirt.org/#/c/7718/1/vdsm/libvirtvm.py
In _getCpuStats (libvirtvm.py) we have this logic to calculate the cpuIdle
stats['cpuIdle'] = max(0.0, 100.0 - stats['cpuUser'])
Problems:
- even with a single CPU, if you fully load the VM's vcpu, the quemu process may take
a little more than 100% CPU time on a multicore host. So in this case 0.0 will be bigger
than -0.1 and vdsm will report no load to engine, this is the rare case
- a more likely situation: create a multicore CPU and run a few parallel dd if=/dev/zero
of=/dev/null. In this case stats['cpuUser'] will be more than 100, and again vdsm
will report no load to engine
Probably this is how it should work:
- since log can be bigger than 100%, we should multiply 100 with the number of vcpus.
- If stats['cpuUser'] > 100 (* nrof(vcpus)), then the load should be the max
possible: 100 (* nrof(vcpus)) not 0.
What do you think?
I'm all for dropping this meaningless number from vm statistics. See my
ack on your
http://gerrit.ovirt.org/7892