Margaret Wasserman <[EMAIL PROTECTED]> wrote:
|But, if we do allocate PI addresses soon, we will need to do that without
|any certainty that we can come up with a scalable PI routing scheme.
Do you really think that there is any question of whether we could if we
actually wanted to? The arguments that it is impossible always come from
people who start with the two assumptions that (a) routing must be done with
the same centralized full-knowledge model that we use now and (b) hardware
absolutely cannot keep up with the mild linear growth of routes in users.
This is an obviously self-fulfilling prophecy, but it has been used to
justify trading linear route growth in users for exponential address usage
in height of provider chain. I think this was a very bad trade to make and
that it will guarantee a perpetual shortage of address space no matter how
many (fixed) bits we start with at the top of the chain.
Previously I proposed (in some detail) a mechanism to retrofit portable
identifiers to IPv6. I suspect that it is isomorphic to some distributed
routing schemes, but keeping the addressing mechanism separate from the
routing mechanism makes it easier to evaluate certain characteristics of
the solution. I'm not claiming that this is the right scheme to implement
(especially given that a patent seems to have been issued for a subset of
the method over a year after I described it) but it does have a property
that makes it worth thinking about, at least in the abstract. Although
there is no obvious way to prove that my method ultimately scales in the
face of increasing renumbering, my method makes no more demand on its
mapping infrastructure than does the DNS. Thus if my method is doomed then
so is the DNS, or, looking at it differently, we would have to solve the
same problem for the DNS to keep it working in the face of renumbering.
|But, I've been told by ISPs who lived
|through the CIDR transition in IPv4 that this _really_ isn't something
|that we want to repeat for IPv6.
I wonder what they meant by that? The CIDR(*) transition was terrible for end
users but ultimately a great boon to ISPs. Recall that we were promised that
CIDR addresses would be _portable_ because the whole exercise was just a
temporary stopgap measure until the hardware caught up. We were supposed to
be able to take our ISP-assigned CIDR block and move it to a different ISP,
with the assumption being that this would happen slowly enough that CIDR would
keep the table growth under control until the hardware caught up. How long
did that promise last? A couple of months? As soon as the economic benefits of
binding addresses to ISPs became obvious all those CIDR blocks became non-
portable and have remained so ever since, massive advances in hardware capacity
notwithstanding. This particular history is part of the reason that I'm so
skeptical about the claims of ready availability of stable global addresses to
end users under v6--economic considerations often trump technical ones.
(*) I use the term CIDR because it is popular, but CIDR itself isn't the
problem. CIDR is just a slightly different way of looking at netmasks.
The problem is the hierarchical address allocation required to make route
aggregation pay off.
|Some folks have argued that easy renumbering would eliminate the need
|for enterprises to have provider-independent addressing, but I don't
|agree. Addresses are stored in many places in the network besides
|the interfaces of routers and hosts, such as access control lists,
|configuration files, .hosts files, DNS configurations, ACL lists, etc.
|In many cases, addresses are stored in nodes on other subnets. So,
|being able to renumber the interfaces of hosts and routers on a
|particular network or subnet doesn't even solve half of the problem.
Those who earnestly argue that easy renumbering would eliminate the need
for PI address space (and please don't confine the need to enterprises--
the stability of my home network is more important to me than that of any
enterprise) are typically talking about a complete solution that would
pervade all the areas that you mention and would require massive interaction
with a variety of tools. IMHO this is not impossible, but those that claim
that it is an easier problem than the PI route problem or even the site local
semantic problem are way off base.
|So, what do we do?
|
|Choices seem to be:
|
| (A) Continue with PA addressing, and accept that enterprises will
| use IPv6 NAT to get provider-independence.
Sufficiently large enterprises will get their own address space by (if
necessary) claiming to be higher-level ISPs.
[...]
|I'm not sure that we'll ever reach anything resembling "IETF consensus" on
|that choice, though.
Given the strong voices that oppose anything that would change the economic
status quo, you are almost certainly correct.
Dan Lanciani
ddl@danlan.*com
--------------------------------------------------------------------
IETF IPng Working Group Mailing List
IPng Home Page: http://playground.sun.com/ipng
FTP archive: ftp://playground.sun.com/pub/ipng
Direct all administrative requests to [EMAIL PROTECTED]
--------------------------------------------------------------------
