I'm amazed anyone in this list would consider PAT in IPv6 as a preferred solution.
If you need fast failover multihoming with consumer-class connections (ignoring the fact that connections will still break when failover occurs), then use NPT between the two providers. Pick your primary ISPs prefix as your internal prefix, NPT to the other one, no port translation needed. Or use ULA, whatever. This multihoming issue could have been solved with better rigor in a few places in the v6 standard, having two routers advertising their two prefixes and default route priorities directly to clients. No complicated multi-wan routers, pizza place just buys two basic consumer routers, plug them both into the same LAN switch, and bam you have two ISPs. Clients get addresses on both prefixes, and default routes to both routers. Router withdraws its default route when it loses the upstream connection, clients deprecate the route, existing connections will fail because there is no longer an upstream link (which will happen with NA(P)T anyway!). Transport protocols see and can use both addresses if they are multihoming aware. Source address selection (rfc6742 rule 5.5) tells clients to pick an address from the higher priority router as the source address, meaning the router advertisement priority is used by the network operator to prefer one ISP over the other. Client software can bind to a specific address if it must use one ISP. Unfortunately for us, rule 5.5 is optional and poorly implemented by OSes, so in reality this does not work that well. If that had been a requirement, it would have solved this issue without any packet mangling, and made transport layer support for multihoming significantly easier. _______________________________________________ NANOG mailing list https://lists.nanog.org/archives/list/[email protected]/message/VKUAQOKO5GNCOMGI5CYVPWK3VRVOPQGZ/
