rigid sphinx
I was playing arround with NVMEoTCP in vmware 7.0u3n and i found that space reclamation does not work or it is not implemented ( this is the equivalent of space allocation for LUN's or SCSI unmap )
I have as environment an hyperflex with 3 nodes running vmware 7u3n and the storage is A250 ontap 9.13p2.
Is there any workaround to this from ontap side ?
frank topaz
not yet implemented: https://kb.netapp.com/onprem/ontap/da/SAN/ONTAP_namespace_space_reclamation_on_NVMe_over_TCP_datastores
(not sure if any storage vendors already supports this)
frosty dome
I think this was implemented in the NVMe standard 2.2 which is from 2021. but until now I do not know any vendor which implemented this
rigid sphinx
oki so we do not overprovision ... but still how do we handle the case when the namespace inside the volume gets full ?
my guess is to use a datastore cluster with multiple namespaces added and let vmware balance the space but even so vmware can not know how "full" is the namespace on ontap side .
main nebula
VMware sees how full the namespace is, because it contains a VMFS filesystem which vmware can control. VMware does not see how full the volume is where the namespace resides in, but if you do not overprovision, that shouldn't be a problem
rigid sphinx
vmware does not see how full is the namepsace that is the problem (* you can see in the attatched document the discrepance between the namespace utilization and the vmare file system )
main nebula
what you're seeing is just thin provisioning of vmdks inside the namespace
don't confuse the layers "vmdk in vmfs on namespace (orLUN)" and "LUN in volume in ONTAP". those are unrelated and each have their own thin-provisioning
rigid sphinx
yes when the namespace will get full this will be disconnected by the storage same as when you provision a lun and have no space allocation on ( that is off by default another strange stuff ) ; and since there is no space reclamation you can not do anything about it ; the host will see the datastore with free space and the namespace will be full ....
main nebula
the "strange reason" that space allocation is off by default is that not all operating systems support it so it's safer to have it off by default