Hi,
I am trying to understand the performance values returned by the
queryMetricsData APIs (same for the VBoxManage command line utility).
>From my testing, the values being returned for a running VM does not seem to
>match the Linux values of CPU usage. The vbox metrics do seem to trend in the
>same direction as the Linux metrics
For testing, I did the following:
1) Wrote utility that read CPU jiffies over fixed delta
2) Use VirtualBox period same as fixed delta in #1 above
3) use prstat and top as validation for the accuracy of the cpu jiffies utility
I used a period of 5 seconds for my tests (but used others with the same
result).
It seems the perf data for a host object matches closely the Linux overall CPU
usage.
The discrepancy seems to be however when comparing virtual machines. The vbox
metrics are lower (by up-to 30% at times) than the linux os level metrics for
the corresponding vm task.
It seems the vbox metrics are stats specific to the 'virtualized cpu'. If that
is the case, it would appear then that the delta between what vbox metrics
reports vs what the linux os level metrics reports for the same VM would be the
VirtualBox 'overhead'.
OK ... so after all that ... essentially ... I am asking if anyone can confirm
that the vbox stats are internal kernel and user usage within the VM based on
the VMs view of available CPU cycles.
Thanks,
Joe
_______________________________________________
vbox-dev mailing list
[email protected]
http://vbox.innotek.de/mailman/listinfo/vbox-dev
