10:51 p.m.
hi,
I am trying to change a behavior in vdsm. When you pass 100% load on a VM, it will stop
reporting further load and will keep telling 100% until the load drops under 100% again in
it's cpuIdle information. This is totally correct if you have only single-cpu
VM's, but it is false when you have multiple vcpu's, I think the cpuIdle
information should not be on a 0-100 scale, but on a 0-100*vcpus scale.
So I submitted this patch to vdsm: http://gerrit.ovirt.org/#/c/7892/2 and Dan pointed out
that some functionality may depend on the value in the 0-100 interval. For me it seems it
is ignored and the load is calculated only from sysCpu + userCpu. Does anyone build on the
cpuIdle value?
Thanks,
Laszlo
9:38 p.m.
----- Original Message -----
Ok.
However, before I abandon this patch: we have a requirement to report cpuSys and cpuUser
separately. Afaik in libvirt cpuUser and cpuSys does not include the actual guest time (at
least not with KVM), and in this way if we only report cpuSys and cpuUser, the sum does
not give the actual load, only a relatively little percentage of it. If we have the
cpuIdle information in engine, we can calculate the guest time. Therefore, should I
- include the guest time in cpuSys or cpuUser?
- add another exported field?
And in both case, we will still have to calculate from cpuIdle because libvirt does not
tell the guest cpu time :-(
Dan.
10:34 p.m.
On Tue, Sep 11, 2012 at 07:38:11AM -0400, Laszlo Hornyak wrote:
Why abandon? I've suggested you to keep it, just make it even simpler.
I am not sure I understand what you are saying, but afaik, libvirt's
relatively-new
http://libvirt.org/html/libvirt-libvirt.html#virDomainGetCPUStats
reports the cpu time spent by the entire qemu process - in guest and
host modes.
Now I'm completely at loss. Why should we calculate cpuIdle per VM?
Haven't we agreed that it is useless?
Dan.
10:52 p.m.
----- Original Message -----
Ok, it is only burocracy, but the new patch will do something completely different than
the original, so it does not seem to make sense to continue this patch. It is more simple
to make another one.
It seems like sysCpu + userCpu < cpuTime, therefore something is missing. I will give
it another try, maybe something wrong with my hosts.
Well, if libvirt exports the guest time in sysCpu, then we do not have to. But it seems it
does not.
Dan.
11:11 p.m.
On Tue, Sep 11, 2012 at 08:52:51AM -0400, Laszlo Hornyak wrote:
Actually, if the patch has the same intentions, I prefer to keep it as
another version - this way it is easier to see if former comments have
been addressed. But it's your code and your call.
I may well be wrong.
http://libvirt.org/html/libvirt-libvirt.html#VIR_DOMAIN_CPU_STATS_CPUTIME
someone would have to compare your report requirement with what libvirt
provides.
Even we should jump throguh hoops to get the data you require, I do not
see how cpuIdle is related (or actually, what it means).
Dan.
4635
days inactive
4637
days old
8 comments
4 participants
participants (4)
-
Alon Bar-Lev -
Dan Kenigsberg -
Laszlo Hornyak -
Omer Frenkel