> in many enterprise/organizational situations, and at least some > home networking ones, communication among hosts on the local > network is very important, perhaps equally important as > communication with outside hosts. I didn't claim otherwise; in fact I suspect that internal communication is often viewed as more important than external communication. However, what I said was "users don't *just* care about local apps" (emphasis added). Which, for instance, is why there is a demand for ALGs in NATs. > Now, in the IPv6 "multiple addresses per host" model, it would > make perfect sense to assign every host on such a subnet an > address that was specific to that subnet (or enterprise) and, > for those hosts that needed external accessibility, an > additional, presumably globally-routable, address. It might seem to make perfect sense at first glance, but I'm not sure that it holds up under analysis. First, as long as hosts are free to assign their own addresses, using DHCP (or RA) as a means of access control to external links doesn't seem very robust. Second, the design decision to use multiple prefixes has consequences for applications that are almost certainly unintended and undesirable. Now having a set of addresses that are stable across provider changes really does seem useful, but I see no inherent reason that such addresses should be only unique (or valid) within the local network. > > Lots of good ideas got passed over in the IPng discussion. > > Despite that, I think that IPv6 ended up "about right" - > > modulo a few warts like A6 and SL and over-reliance on address > > selection. By "about right" I mean that I think that if the > > warts are fixed it will become feasible to gracefully extend > > the IPv6 architecture in useful ways - such as to provide the > > capability to use identifiers rather than locators - without > > invalidating hosts or apps that were written to the basic IPv6 > > model. > > Keith, I think I agree with you about the "despite that" part of > the above (I certainly agree about the "passed over" part). > However, if our confusion about how to best handle multiple > addresses per host, and where to do so, results in our moving to > a "one host, one address" model and a flat, IP-level, address > space, I think some of the "graceful extension" potential will > disappear. Indeed, if we go down that path --which I think you > have been advocating, although I'm not sure-- I suspect that we > had better have a general routing solution that is not dependent > on address aggregation before we move much further forward with > IPv6 deployment. Frankly I don't know where this needs end up. I don't think it ends up with a purely "one host, one address" model, for a variety of reasons. I do think it might end up with a model where there are "distinguished addresses" (or if you prefer, identifiers) for hosts that are generally preferred over other addresses or identifiers, but can be automatically forwarded and/or mapped to locators that can be routed by traditional means. I'm looking at various ways of doing this, some of which involve site-relative locators and some of which don't. So far I haven't seen a case wehre relative locators are required. They can make certain cases easier- particularly isolated newtorks. But they seem to make the overall solution more complex and less robust, even when they're not exposed to applications. I'm still working on understanding this landscape, so I might find a compelling case for site-relative locators. But I suspect not. Keith