I have a network consisting of two MX480 and several QFX5100 VCs (see attached picture). We evaluated MC-LAG for a while but this was not as flexible and stable as we needed. So we switched to EVPN-MPLS with all-active multihoming. We have now done some tests and everything is working as expected. I have some questions in regards to scaling the whole platform.
1) Scaling and difference between VLAN-Aware Bundle and VLAN-Based Service We have chosen to use VLAN-Aware Bundle service. With this every customer has its own VLAN terminated on his own irb interface. Would it be better to choose VLAN-Based service so can configure a single EVPN instance for every VLAN? Are there any limititations? I am thinking of creating a EVPN instance for every single customer (vlan). But I think that is too much of it.
2) We would like to provide some of our customers with BGP
How could we achieve this in our current setup? Using the irbs is not really possible in all-active multihoming. With MC-LAG we used VRRP, so we could use two different IPs on both routers for BGP sessions with a customer. I have thought of creating a logical system on each of the routers and using a separate VLAN-Based EVPN Service just for BGP customer termination. Is this too much? Maybe there is a simpler way of doing this.
Attached you find a network diagram. The routers are latest MX480 with Dual RE, running on 17.4R2. The core switches are QFX5100 running on latest recommended version.
Vlan-Aware Bundle service scales a lot better than the Vlan-Based service becuase the scale depends on the ammount of Bridge-Domain and MAC adresses. If you use the Vlan-Based service, the ammount of VRFs supported by the platform also plays a role.
Regarding running routing protocols on IRB, you should use "Virtual-Gateway-Address" concept (instead of anycast MAC/IP) where each Gateway can share a common IP/MAC (same as VRRP) and at the same time have it's own/dedicated IP and MAC.