Open Stack

On Wed, 2023-01-11 at 23:10 +0100, Ignazio Cassano wrote:
> Hello All,
> I would like to know if with numa in kvm hypervisor I must disable merge
> across nodes ? If yes, is there any metadata to do it ?
no nova has numa toplogy awareness for vms with a virutal numa toplogy

i.e. if you want vms ot have numa affinity you can request a numa toplogy explcitly by using hw:numa_nodes=<number of virutal numa nodes for the vm>
it can also be done via the image with hw_numa_nodes.

nova will also generate a request for a singel numa node if  you use specific feature like cpu pinnign or hugepages  to name two cases. i belive pmem
also does it so that is not ment to be an exaustive list.

all vms that have an implciet or explict numa toplogy should have hw:mem_page_size or hw_mem_page_size set to an allowed value
(small,large,any,<intager_page_size[KMG]>) for example hw:mem_page_size=small will create a singel numa node guest using implcit numa request as a
result of specifying
a page size and small will ensure that hugepages are not used for the vm. on must cpu archtecutres small means the default 4k pages are uesed.
such a vm will be pinned to a spcific host numa ndoe but will float over cpu on that numa node sicne cpu pinnign is not enabeld.

if the vm does not request a numa topology we will not do any numa aware placment or affinity enforcement.

nova does not support mixing numa and non numa vms on the same host and its left to the operator to ensure that useing any of the means supproted
by placement or schduler filters.

> Ignazio