As I understand, you are counting total stolen milliseconds.
At this point you know few about the steal impact.
If you calculate the stolen milliseconds divided by the amount of VMs you'll get the stolen milliseconds per VM.
Then, dividing the stolen milliseconds by the actual milliseconds on which the stolen value applies (the interval of counting the stolen milliseconds), you'll get the % of time that a VM has been stolen (in average).
So the calculation would be % stolen per VM = stolenms / #VMsOn / intervalms