[PVE-User] High CPU utilisation on KVM VM1 affects VM2

[Martin Schuchmann]

Hi there,

I had never tried to operate with the CPU-Units option for KVM guests, 
but yet read the manuals and the older forum entries.

Today we experienced the following problem:

Host:

16x vCPU Xeon E5520 (2x 4Core Xeon => 8 real cores), 64GB RAM

pveperf
CPU BOGOMIPS:      72529.52
REGEX/SECOND:      821793
HD SIZE:           18.33 GB (/dev/mapper/promo3-root)
BUFFERED READS:    417.91 MB/sec
AVERAGE SEEK TIME: 4.48 ms
FSYNCS/SECOND:     1913.66

vzcpucheck
Current CPU utilization: 3000
Power of the node: 1813240

KVM VM1:  4x CPU / 8 GB RAM / CPUUNITS 1000 / W2008 R2
KVM VM2:  6x CPU / 8 GB RAM / CPUUNITS 1000 / W2008 R2
OpenVZ CT1:  1x CPU / 512MB RAM / CPUUNITS 1000
OpenVZ CT2 : 1x CPU / 512MB RAM / CPUUNITS 1000

Some software on  VM1 failed, and pushed up the CPU utilisation to 120% 
(!?). Working with VM1 was not possible anymore. We tested this 
behaviour for a few times with the same result - as soon as the program 
was started the serverload rises in some extreme manner - as we tried 
other programs  everything was kept fine.

But also VM2 got a very high cpu utilisation, too (at those moments ) 
and at one situation VM2 went away with a bluescreen. During the time of 
high utilisation by VM1  also CT1 and CT2 did show the same behavior on 
CPU: normally they stay at about nothing - but in those moments they 
rised their cpu utilisation - from 0,01 to 2%.

Is there any possibility to avoid this negative influence on other 
guests? Setting kind of a limit?

Since today I thought, if I do not allow the usage of all of the real 
cores by only one Guest, I won't ran into such a situation - but 
obviously this idea doesn't work.

"man qm" tells according to "-cpulimit":         NOTE: This option is 
currently ignored.

On "-cpuunits" there is mentiond a maximum of 500.000 - would the 
solution be something like:

KVM VM1:  cpuunits 250.000
KVM VM2:  cpuunits 200.000
OpenVZ CT1:  cpuunits 50.000
OpenVZ CT2 cpuunits 50.000

?

Or would it be the right way setting "-cpuunits" to 0 (to disable the 
fair-scheduler) and so working with the accounting of cores  for each 
guest only?

Does anyone have some experiences or strategies with that?

Thank you in advance!

Martin