Create multiple igroups, each one with maybe a single initiator (KVM host)
Modify pdisk logic, so it knows in advance which KVM host the LUN should be assigned to (gathered from ON)
Map the LUN ONLY to the igroup containing the initiator of the corresponding KVM.
If we use this approach:
There should be a lesser number of LUNS seen by each KVM, which could improve performance and mitigate the overhead associated to LUN rescanning on KVMs.
We do not have to apply the model of binding LUNs to several igroups, which is possible, but not generally recommended
We work out the upper limit of 256 hosts, as the igroup limit is much higher.
HA considerations (VMs movements between hosts) are not on the scope, although could be nice to foresee future modifications that would be required in this model.
Proposal:
Create multiple igroups, each one with maybe a single initiator (KVM host) Modify pdisk logic, so it knows in advance which KVM host the LUN should be assigned to (gathered from ON) Map the LUN ONLY to the igroup containing the initiator of the corresponding KVM.
If we use this approach:
HA considerations (VMs movements between hosts) are not on the scope, although could be nice to foresee future modifications that would be required in this model.