On Thu, May 6, 2010 at 7:16 PM, Dean, David (I/S) <[email protected]> wrote: > Is there a way to see what CP is doing during “steals”?
Yes, VM monitor data provides metrics to determine that. Linux itself cannot tell why, and should not want to. In many cases it's due to CPU contention on z/VM or LPAR level. In some cases it is CP overhead. And in some cases it's just incorrect configuration of the virtual machine. When during some period there is more demand for CPU resources than what you have available, some virtual machines will have to wait to get the cycles they wanted. Which virtual machine needs to suffer most is decided by the z/VM performance analyst who translate business objectives into configuration and tuning parameters. When people complain about high "steal rate" you should probably ask whether they still meet their service levels... The fact that someone sometimes has to wait is not the deciding factor. The tuning challenge is to make things work with less than infinite resources. PS I think the term "steal" is a wrong name. After all, how could a Linux virtual machine claim that CPU cycles were stolen when it never had them (and will actually get them a bit later). I would rather call it "involuntary wait" or so. Rob -- Rob van der Heij Velocity Software http://www.velocitysoftware.com/
