On Wed, Sep 6, 2017 at 2:47 PM, Gianluca Cecchi <gianluca.cecchi@gmail.com> wrote:
Hello,
I was talking with a guy expert in VMware and discussing performance of VMs in respect of virtual cpus assigned to them in relation with mapping with the real hw of the hypervisor underneath.

One of the topics was numa usage and its overheads in case of a "too" big VM, in terms of both number of vcpus and memory amount.
Eg: 
suppose host has 2 intel based sockets, with 6 cores and HT enabled and has 96Gb of ram (distributed 48+48 between the 2 processors)
suppose I configure a VM with 16 vcpus (2:4:2): would be the mapping respected at physical level or only a sort of "hint" for the hypervisor?
Can I say that it would perform better if I configure it 12 vcpus and mapping 1:6:2, because it can stay all inside one cpu?

Hard to say without relationship to the workload. You are losing 4 vCPUs - perhaps those can be used for something (the OS) while the rest of them could be used by the application, for example?
 

And what if I define a VM with 52Gb of ram? Can I say that it would perform in general better if I try to get it all in one cpu related memory slots (eg not more than 48Gb in my example)?

Hard to say without relationship to the workload - will it need all the memory? Will it be accessing all of it, in random order?

If you've maxed out one node, you just need more memory from the other node. 
 
Are there any documents going more deeply in these sort of considerations?


It is so workload dependent that there will not be a one size fit all.

Also, if one goes and sizes so that the biggest VM is able to all-stay inside one cpu-memory, does it make sense to say that it will perform better in this scenario a cluster composed by 4 nodes, each one with 1 socket and 48Gb of memory instead of a cluster of 2 nodes, each one with 2 sockets and 96Gb of ram?

You could have used affinity.

See [1] for some details on Redis. Note that IO (specifically network) is just as important - and its impact is much more profound.
Y.

[1] https://redis.io/topics/benchmarks
 

Hope I have clarified my questions/doubts.


Thanks in advance for any insight,
Gianluca  

_______________________________________________
Users mailing list
Users@ovirt.org
http://lists.ovirt.org/mailman/listinfo/users