On (2014-02-11 09:27 -0500), Thomas Nadeau wrote: > I think you are missing the point that was raised: its not that BGP > will not scale well; as a protocol its clearly shown that it can ship around > millions of routes and lots of other stuff as it has become the "dump truck" > of the protocol world. The question is more of the approach - do we want to > be shuffling around /32s (regardless of protocol) given that there are other > solutions that work well without doing so? I would also question the use of > BGP in a Data Center given past operational feedback in NVO3 for example, but > that is orthogonal to the question of the mechanics of this solution.
One example why I think people might want L3 (i.e. shuffling around /32s) Imagine you have in your own network DC in San Francisco and New York. Subnet is advertised from San Fran. Client in New York complains high latency, so you use XLAN or what not to migrate their server to New York, to be closer to customer. Before: client -> ny -> sf -> ny -> client After: client -> ny -> sf -> ny -> sf -> ny -> client You moved server closer to customer and you doubled your latencies and reduced MTBF. It seems quite simple, vswitch runs BGP to advertise /32 hosts, uses proxy-arp to connect disconnected hosts in subnet. During convergence GRE tunnel to bounce wrongly routed packets to correct vswitch, hitlesss L3 migration with optimal routing. Why would you not want L3? -- ++ytti
