Joe Abley wrote:
On 19-Feb-2007, at 18:06, Alastair Johnson wrote:
Interestingly, though, finding operators who can describe what is wrong with it beyond "that's not how we were taught things work" is quite difficult.
Well, as I recall, the idea behind using multiple provider address blocks for multihoming was that you would use dynamic configuration to figure out where you were in the world, and A6/DNAME DNS records to get services back to where you wanted to them go. A6 and DNAME have long since been consigned to the "too hard" basket. I think the problem is simply that no-one has really convincingly demonstrated that multihoming on different blocks is "nice", especially with DNS caching, connection caching etc. Multihoming on multiple addresses pushes a lot of the smarts onto the client (as opposed to doing it in the network), which rather increases the scope for things to go wrong. The thing is that IPv6 was seen as a chance to do everything over. Most of the promises have either fallen by the wayside or been layered on top of IPv4. The only thing that IPv6 really promises that can't be delivered with IPv4 is long addresses. Depressingly, the TUBA proposal (TCP and UDP over Big Addresses), which would have just replaced IP datagrams with OSI connectionless service datagrams, was dropped (along with a couple of other proposals) in favour of what became IPv6, OSI CLNS was widely implemented (albeit rarely deployed) at the time, and while the proposal needed some work, it could have delivered the needed long addresses much, much sooner. But of course religious wars, NIH and second system syndrome together put paid to that idea. (That and a complete disconnect between the people defining the protocol and the people who knew how to make routers switch packets fast; the former simply failed to believe that CLNS variable length addresses could be switched as fast as fixed length ones, while the latter knew damn well that they could.) -- don