At Mon, 29 Apr 2002 12:54:51 -0700, Steve Deering wrote:
> 
> No more so than the use of DHCP relays does.  We're talking about the
> IGP within a site, and taking advantage of what it already can do.

Nope, you're storing new information in the IGP that wasn't there
before.  But back up and take it one step at a time.

In the abstract, DHCP is a multicast query, unicast response protocol.
A client yells "yo, anybody got a foo server?" and gets back zero or
more responses, from which it picks whichever response(s) it likes
(yes, I'm oversimplifying slightly, and am completely skipping over
stateful address assignment, which isn't relevant here, but in
principle that's how post-address-config works using DHCP).  Relay
agents (again ignoring stateful address assignment) are basicly a tool
for getting DHCP to work in environments that (for whatever reason)
don't suport multicast in its fully glory.

The key difference between the DHCP model and the anycast model is
that in the case of the DHCP model it's at least possible to set
things up so that the entity that's telling you where the DNS server
is actually -is- the same entity as the DNS server.  Thus, if the DNS
servers you're using have stopped responding, you can re-ask "where's
there a DNS server?"  and find out who's available now.  Similarly,
you can keep track of which addresses you've been talking to and how
well they worked (a DNS server can be up and running and still be
useless due to misconfiguration, bugs, system load, network load,
acts of malice, ...).

In the anycast model, you're at the mercy of the system that's
maintaining the anycast route: you can't query directly to find out
where the DNS server is, you have to wait for the routing system to
figure it out, and the routing system doesn't tell you when it change
the binding between a particular anycast address and the server behind
it, so you can't usefully keep track of which DNS servers have been
misbehaving recently.

So I stand by my previous statement that the anycast proposal moves
data and functionality from the edge into the core of the network.

I suspect that further discussion on this particular point would be
more suitable for a beer BOF in Yokohama than for this mailing list.

> (Now that I mention it, perhaps DHCP relays ought also to be replaced
> by the use of a well-known, site-local anycast address.)

That'd be another step towards centralization, since it assumes an
environment in which there's only one[*] DHCP server that your client
might want to talk to, and completely loses the capability to send the
same question to whatever DHCP servers are in range.

[*] Or two, or three, or whatever small fixed number of anycast
addresses you choose to allocate, and to date I haven't seen an
explanation of how servers anycast::1, anycast::2, and anycast::3
are supposed to sort out which one gets which anycast address without
some kind of coordination mechanism that either involves yet another
new protocol or something that quacks like manual configuration.
--------------------------------------------------------------------
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]
--------------------------------------------------------------------

Reply via email to