1. neutron knows which node each port is bounded in and establishes vxlan tunnel among the compute nodes via ovs vxlan type port with remote vtep ip designated

2. no. one pair of path port is enough

3. ovs flow table does it

4. yes. but commonly 4096 is enough for one compute node even if each port bounded with the compute node has different vlan tag.

Regarding Q1, the first packet would normally be an ARP broadcast, which would be flooded to all compute nodes, causing MAC-learning of the source on the way forwarding. The ARP-response would then cause MAC-learning for the original requested destination, and voila! :-)

With l2pop I dont know the actual details but since the whole idea is to pre-populate the FDB of the OVS, I have simply assumed that OVSDB or ovs-cli-tools are used to populate the FDB of each OVS on each compute node.

I cant find the reference now, but I think I've read somewhere that the OVS tag in bridges, is not a 802.1Q tag, meaning its 24..32 bits or something like that, and doesnt have the limitation of max ~4000 vlans.