Isolating a processor core

[TheVMinator]

Is it possible to dedicate an entire processor core to a virtual machine?  With processor affinity, I dedicate a virtual machine to a processor core.  That VM can only run on the core that it has affinity with.  However, other vms can still run on that core.  How can I isolate an entire processor core to be used by one and only one virtual machine?  The other processor cores should be free to run any of a large number of virtual machines - so using processor affinity for all vms so that one vm can have its own core is not an option.  I just need to dedicate one core to one vm and keep every other vm from accessing it.

(Note - this is all terrible design and not best practice which I realize - it is beyond my control to say how done but need to know if it is theoretically possible)

Thanks for input

[frankdenneman]

You can, by using CPU-affinity.

But don't do it, cause it will likely offer performance decrease than increase if thats your main goal.

Read this article and you know which elements to configure.

http://frankdenneman.nl/2011/01/beating-a-dead-horse-using-cpu-affinity/

Frank Denneman

co author of: vSphere 4.1 HA and DRS Technical Deepdive - out now on Amazon

[TheVMinator]

Great article - and from what I see the only way to isolate a core to one and only one virtual machine is to set CPU affinity on EVERY OTHER vm on that host - even everything that can be done with reservations and shares will not necessarily create that result.

[frankdenneman]

You are correct. But do not forget every vCPU will be scheduled according to the rules of the CPU scheduler.

That means:

• 1 vCPU per pCPU
• Timeslicing for fairness principle, even if you isolate a pCPU. so a vCPU will be descheduled if it reaches the end of its timeslice, because no other vCPU run on the pCPU it will be scheduled imediately.
• Other worlds, BHs and interrupts need to be scheduled too, so assign a second pCPU to eliminate vCPU and worlds competing for time.
• The VMkernel uses all available pCPUs for scheduling its processes, it does not take a VM with CPU-affinity into account.

Tip:

• If you are running on a NUMA node, set memory affinity to avoid remote memory.
• "De-assign" the second pCPU on all other virtual machines as well, because the CPU-scheduler can (most unlike, but there is no guarantee) migrate the vCPU thread to the other processor for load-balancing purposes.

Frank Denneman

co author of: vSphere 4.1 HA and DRS Technical Deepdive - out now on Amazon

[TheVMinator]

OK thanks again