Problem That Additional Paths Can Solve
BGP routers and route reflectors (RRs) propagate only their best path over their sessions. The advertisement of a prefix replaces the previous announcement of that prefix (this behavior is known as an implicit withdraw). The implicit withdraw can achieve better scaling, but at the cost of path diversity.
Path hiding can prevent efficient use of BGP multipath, prevent hitless planned maintenance, and can lead to MED oscillations and suboptimal hot-potato routing. Upon nexthop failures, path hiding also inhibits fast and local recovery because the network has to wait for BGP control plane convergence to restore traffic. The BGP Additional Paths feature provides a generic way of offering path diversity; the Best External or Best Internal features offer path diversity only in limited scenarios.
The BGP Additional Paths feature provides a way for multiple paths for the same prefix to be advertised without the new paths implicitly replacing the previous paths. Thus, path diversity is achieved instead of path hiding.
This section describes in more detail how path hiding can occur. In the following figure, we have prefix p with paths p1 and p2 advertised from BR1 and BR4 to RR1. RR1 selects the best path of the two and then advertises to PE only p1.
In the figure above, we also see prefix x with path x1 being advertised from BR2 to BR3 (which has path x2) with local preference 100. BR3 also has path x2, but due to routing policy, BR3 will advertise to the RRs x1 (not shown) instead of x2, and x2 will be suppressed. A user could enable the advertisement of best external on BR3 and thereby advertise x2 to the RRs, but, again, the RRs advertise only the best path.
Suboptimal Hot-Potato Routing Scenario
In order to minimize internal transport costs, transit ISPs try to forward packets to the closest exit point (according to Interior Gateway Protocol [IGP] cost). This behavior is known as hot-potato routing. In the distributed RR cluster model of the figure below, assume traffic coming from LA must go to Mexico. All links have the same IGP cost. If there are two exit points toward Mexico—one toward Austin and one toward Atlanta—the border router will try to send traffic to Austin based on the lower IGP cost from LA toward Austin than toward Atlanta. In a centralized RR model where the central RR resides where RR3 is (and RR1, RR2, RR4, and RR5 do not exist), the closest exit point toward Mexico, as seen from RR3, might be Atlanta. Sending the traffic from LA toward Atlanta results in suboptimal hot-potato routing, which is not desirable.
In Dynamic Multipoint Virtual Private Network (DMVPN) deployments, BGP is being used for scaling. In the figure below, Z is connected to both spokes S6 (NY) and S7 (Boston). The S7 links to the hubs have lower IGP costs than the S6 links to the hubs. There are physical links not shown that connect S5 to S6 and S6 to S7, with IGP costs lower than those to the hubs. Spokes S6 and S7 will send an update to both hubs H1 (Chicago) and H2 (Detroit). The RR hubs will then select the best path based on their lower IGP cost, which might be S7. The spoke S5 (Raleigh) will receive two updates from the RRs for Z with S7 being the next hop, even though, in this scenario, it might be preferable to pick S6 (NY) as the next hop.