Deepak Jain wrote: | Best guess approximation? Say not more than 1 minute per router (* N | router hops between 1 edge of network and the other?) typically more like | 10 seconds on moderately loaded routers/links? Hm, where does 1 minute per router come from? The only interesting per-router issue is that there may be multiple congested gateways between the source of an iBGP update and some of its destinations. Since iBGP rides on top of TCP it is sensitive to dropped segments, and where traffic pours into a forwarding loop, this can cause multiple full RTOs. Hierarchicalizing the iBGP mesh is a win in part because it localizes some TCP connections such that there are fewer (or no) intermediate gateways that may be congested, thus reducing the liklihood of dropped segments. Likewise, prioritizing iBGP TCP segments over other traffic also reduces the incidence of drops. As far as I know, all large networks currently do both of these things. Finally, fast convergence is a false deity: increasing control traffic in the face of the instability that is endemic to the Internet is a recipe for disaster. Staying slow and being stable is a more intelligent engineering goal; the drawback is that some routing protocols create transient forwarding loops whose lifetimes can be extended under load -- and the load may stem from forwarding packets in a fast loop. A permanently non-converging iBGP mesh replete with loops and black-holes is not just a theoretical possibility. :( Sean.