Yeah generally if usage is lower than demand you have 1 of 2 issues.
1) The VM simply needs more resources because its been limted or needs more than the amount of CPUs allocatted
2) There is real contention and just cant get what it needs
Can I have your thoughts on the attached images which I took at cluster and host level?
Is this actually an issue? As you can see the demand is higher than usage but plenty of capacity left on this cluster. The workload indicator seems to detect no issues.
When I check the hosts in the cluster I see demand higher than usage but little difference (i.e 43 Ghz Capacity, 12 Ghz Demand, 11 Ghz Usage)
and when I look at the individual VM's I see something similar. Demand is only a little bit more than usage in most cases. How much difference is considered bad?
Would you actively try to get the demand and usage values closer in this scenario? Should I ignore the difference at cluster level?
I did check the CPU | Contention metric at the cluster level and it peaked around 4.6% in the last week but averages around 2%
For those images I would say everything looks ok, the demand and usage is a bit far removed when looking at the cluster level and sometimes host level if they are large. and if there are limits set or large SMP VMs.
There is no real rule on what difference is bad. its all relative.
Its good to think about it as being reactive a VM wants resources and the hypervisor reacts to that. usage always following demand, it will be very rare even with zero overcommit that demand and usage are equal.usage will generally trail demand by a small amount.
Its best to use this in conjunction with CPU ready and contention and a bit of co-stop. to determine if there is an issue.
Its worth noting CPU contention though is directly related to latency, and latency is heavily affected by CPU state. meaning if power mode is balanced or anything else other than high performance contention can be artificially increased with the change in CPU states
Having a look at the power mode, it is set to balanced as per default. Is it in general recommended to use high performance?
The cluster in question is for general production purposes and power savings (cost wise) is no concern.
Yeah I generally always recommend High Performance on hosts unless power is a real concern.