I think I misunderstood the problem. I recently saw that vms cannot access the other vms 10.0... addresses in the same cluster if the source and target are on different compute nodes. I think that was the real source of "randomness" I was seeing. The floating address was probably a red herring. This was using multi-host and running nova-network on each compute node. Is this a bug? If not, how are a set of vms in the cluster supposed to communicate?
I think I misunderstood the problem. I recently saw that vms cannot access the other vms 10.0... addresses in the same cluster if the source and target are on different compute nodes. I think that was the real source of "randomness" I was seeing. The floating address was probably a red herring. This was using multi-host and running nova-network on each compute node. Is this a bug? If not, how are a set of vms in the cluster supposed to communicate?