On Wed, Dec 31, 2008 at 9:06 AM, Harder, Pieter <[email protected]> wrote:
> So, would this be right? > > Linux steal time = processing time for other VMBKs + non-SIE processing for > own VMBK > > Non-SIE processing for own VMBK = ( TV ratio - 1 ) * SIE emulation time for > own VMBK I have to be careful in phrasing this, or mud will be thrown at me again from the East (and with current temperature that is painful) :-) Clearly the CP portion of virtual machine CPU usage is reported in the Linux steal time. The other component is the time where the virtual CPU is waiting to be dispatched and Linux is aware of it. We have some metrics about queue sampling, but that does not tell me about Linux. So "waiting for CPU" breaks down into a portion that is within the steal time, and a portion that is not measured. And there is a component where PR/SM attributes time to the VMDBK (analogous to the ttime-vtime portion for CP). Because of design decisions in Linux, this is not measured either. In normal steady workload situations, both not-measured portions appear to be fairly small and constant. My view is biased because I am more often presented data from situations where things do not behave as expected. So as long as you have both the Linux and VM metrics, you can determine which part of steal time was for CP services on behalf of the virtual machine, and which part was because of contention with other virtual machines. Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/ ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to [email protected] with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390
