Reflect:
client -> clients and non-clients
non-client -> clients
No Reflect:
non-client -> no-client (normal ibgp)
always advertises to eBGP peers (normal ebgp)
eBGP learned prefixes, advertised to client and non-clients (normal ebgp)
Full Mesh iBGP:
- between RRs
- RRs and non-clients
- clients just need iBGP to RRs.
Reflects ONLY the best route
RRs dont modify on reflected routes: NEXT_HOP, AS_PATH, LP and MED.
Prevent routing information loops: ORIGINATOR_ID and CLUSTER_LIST
Clustering:
- ORIGINATOR_ID: The first RR creates the Originator_ID and sets it to the BGP router ID of the router that originates the route. So when a client receives a route with its own Originator ID, it is dropped.
- CLUSTER_LIST: if the local CLUSTER_ID is found in the list, the route is discarded.
*This is done ONLY in RRs.
This is ONLY created or updated on a RR during Reflection.
Hierarchical Route Reflection: 2 levels: - level1 RRs are clients of level2 RR -> level1 RR dont need full mesh between them - level2 RRs are full mesh between them. Consider: size of top-level mesh number of alternative paths
Hierarchical Route Reflection:
2 levels:
– level1 RRs are clients of level2 RR -> level1 RR dont need full mesh between them
– level2 RRs are full mesh between them.
Consider:
- size of top-level mesh
- number of alternative paths
RR Design Priciples:
- keep logical and physical topologies congruent to increase redundancy and path optimization and prevent loops
— follow physical topology
— change physical topology
— modify logical topology
— follow physical topology
— session between RR and non-client shouldnt traverse a client
— session between RR and client shouldnt traverse a non-client
- use comparable metrics in route selection to avoid convergence oscillations
(ie MED – only used between prefixes from same neighbor AS – not used for different AS)
— full ibgp mess –> no
— always-compare-med –> no
— deterministic-med -> ok
— med=0 (via RM in) -> ok
— bgp communities ->
- set proper intra and inter cluster IGP metrics to avoid convergence oscillations
— multicluster RR architecture: intracluster metrics lower than intercluster.
- modify next-hop with care. Do so only to bring the RRs into forwarding path.
- use peer groups with RR to reduce convergence time.