Performance impact of enabling Memory (not CPU) hot plug?

[mikeely]

Looking around online, the questions surrounding CPU hot add seem very well answered. What I haven't been able to find is a similar discussion on enabling memory hot plug.

Here are the basic questions I'd like to know answers to:

1. Do we see the same concerns around cross-NUMA communication latency with RAM as we do with CPU?
2. Does enabling memory hot plug disable vNUMA even if CPU hot add is not enabled?
3. Is there some other overhead when this feature is enabled that would be absent with it disabled?

Thanks in advance.

[mikeely]

(bump)

[homerzzz]

1. You do not really need to be concerned with NUMA imbalance or remote memory latency if you are running vSphere 6 and later.

2. No

3. Not that I know of. Enabling this feature across the board is not recommended as best practice.

[mikeely]

Enabling this feature across the board is not recommended as best practice.

Can you expand on this a bit? I'm wondering what the detrimental effect of doing this might be that such a recommendation is made.

[homerzzz]

My thoughts on this VMware recommendation is based on planning and right sizing vs. no planning and relying on hot plug. If you know the resource requirements prior to deployment based on data, you are able to make sure the underlying server hardware, network and storage can support the VM requirements. Deploying VMs with a standard template with hot plug enabled makes it more difficult to plan.

That being said, I do enable memory hot plug on many VMs due to the fact that getting proper requirements seems impossible at times. Customers think bigger is better! I guess it really depends on the environment. Maybe in a self serve environment, instead of having a small, medium and large VM choice, just have a small choice that can automatically add memory if needed. You would certainly need to monitor capacity.

Not sure if any of this really answers your question...