Running openstack-nova-compute in a horizontally scaled out fashion (e.g. two instances pointing to one ESXi cluster via the vCenter driver) is not currently recommended and there have been issues with this e.g.:

• https://bugs.launchpad.net/nova/+bug/...

An alternative in the mean time could be to run the openstack-nova-compute service on the ESXi infrastructure itself in a single VM and use vCenter's HA support to keep it up. This does introduce a certain amount of a chicken and egg situation (though arguably no different to the KVM driver where the compute service also runs on the host it is managing) but avoids the potential issues currently inherent in running multiple instances of the service pointing at a single ESXi cluster (I say potential as there is limited testing in this area).

It depends on the hypervisor type. For VMWare you will need the vCenter to enable HA and will handle the High Availability similar to XenServer. However for the KVM Based Hypervisor, you will need to incorporate the use of the host-evacuate api from openstack.