[ovirt-users] Re: VM CPU Topology
nelson.lameiras(a)lyra-network.com writes:
Hello,
We are currently running oVirt 4.3.10
Our oVirt hypervisors (HV) have 2 cpu sockets * 6 cores * HT = 24 vcpu
Our VMs (centos7) range globally from 2vcpu to 8vcpu
oVirt allows to configure - per VM - the following 3 advanced
parameters : virtual_sockets : cores_per_virtual_socket :
threads_per_core
We make sure that threads per core is always 1 (so no question there)
But, for the other 2 parameters, we are unsure of the correct
configuration, and if there is a performance penalty on bad
configuration.
Let's consider a 4vcpu VM
1- Is there a performance difference betwenn 1:4:1 and 4:1:1 configuration ?
2- When should we opt for one or another configuration ?
2- Our VMs total CPU provisionning sum is twice the hypervisors
capacity, but they are mostly idle so it is not an issue, but can this
influence configuration choice above ?
I'm not sure anybody knows an ultimate answer and/or a simple rule of
thumb. My guess would be that generally there should be no significant
difference. But there may be application specific considerations,
especially when NUMA or something similar is involved.
I think the best what can be done is to test both the configurations
with your particular applications and setup and see if any of them
provides a systematically better performance than the other one.
Regards,
Milan
Thank you for any information that can enlighthen us, since we are
worried that we are suffering from bad performance due to naive cpu
configuration choices.