Research: BGP Routers and Parrots
The BGP specification suggests implementations should have three tables: the adj-rib-in, the loc-rib, and the adj-rib-out. The first of these three tables should contain the routes (NLRIs and attributes) transmitted by each of the speaker’s peers. The second table should contain the calculated best paths; these are the routes that will be (or are) installed in the local routing table and used to build a forwarding table. The third table contains the routes which have been sent to each peering speaker. Why three tables? Routing protocols standards are (sometimes—not always) written to provide the maximum clarity to how the protocol works to someone who is writing an implementation. Not every table or process described in the specification is implemented, or implemented the way it is described.
What happens when you implement things in a different way than the specification describes? In the case of BGP and the three RIBs, you can get duplicated BGP updates. What do parrots and BGP have in common describes two situations where the lack of a adj-rib-out can cause duplicate BGP updates to be sent.
The authors of this paper begin by observing BGP updates from a full feed off the default free zone. The configuration of the network, however, is designed to provide not only the feed from a BGP speaker, but also the routes received by a BGP speaker, as shown in the illustration below.
In this figure, all the labeled routers are in separate BGP autonomous systems, and the links represent physical connections as well as eBGP sessions. The three BGP updates received by D are stored in three different logs which are time stamped so they can be correlated. The researchers found two instances where duplicate BGP updates were received at D.
In the first case, the best path at C switches between A and B because of the Multiple Exit Discriminator (MED), but the remainder of the update remains the same. C, however, strips the MED before transmitting the route to D, so D simply sees what appears to be duplicate updates. In the second case, the next hop changes because of an implicit withdraw based on a route change for the previous best path. For instance, C might choose A as the best path, but then A implicitly withdraws its path, leaving the path through B as the best. When this occurs, C recalculates the best path and sends it to D; since the next hop is stripped when C advertises the new route to D, this appears to be a duplicate at D.
In both of these cases, if C had an adj-rib-out, it would find the duplicate advertisement and squash it. However, since C has no record of what it has sent to D in the past, it must send information about all local best path changes to D. While this might seem like a trivial amount of processing, these additional updates can add enough load during link flap situations to make a material difference in processor utilization or speed of convergence.
Why do implementors decide not to include an adj-rib-out in their implementations, or why, when one is provided, do operators disable the adj-rib-out? Primarily because the adj-rib-out consumes local memory; it is cheaper to push the work to a peer than it is to keep local state that might only rarely be used. This is a classic case of reducing the complexity of the local implementation by pushing additional state (and hence complexity) into the overall system. The authors of the paper suggest a better balance might be achieved if implementations kept a small cache of the most recent updates transmitted to an adjacent speaker; this would allow the implementation to reduce memory usage, while also allowing it to prevent repeating recent updates.