RE: Newbie Question about addressing impacts

["Fleischman, Eric" <eric.fleischman@boeing.com>]

Friends,

I am struggling with these issues and don't feel that I have a solution. The view from my knot hole is that very large corporate networks (i.e., Fortune 100, governments) are likely to refuse to renumber or support multiple internal addresses due to cost reasons (e.g., cost of renumbering itself, explosion of routing table entries for IGP, DNS issues, directory issues, DHCP confusion, QoS confusion (e.g., for policing and marking), etc.). Thus, huge corporations are likely to demand that we own our own addresses and that ISPs would therefore have less aggregation as a result. Even should "the system" not permit this, it shouldn't be hard to find ISPs willing to support us in the manner we require for "the right price." Thus, it is inevitable (in my own mind) that this is what the Fortune 100 and governments will do and the IETF really has no say in the matter -- unless it can come up with a transparent approach to remove the penalty of renumbering from our operations. That is, should the penalty of readdressing be removed, then we wouldn't be compelled to "do our own thing" in this matter.

On the other hand, such behavior is unacceptable from a "Internet citizenship" perspective and does not address the similar business issues that moderate size corporations (e.g., 100,000 employees) have, which would influence them to behave in a manner similar to the Fortune 100. But if moderate size corporations do this, then why not smallish corporations (e.g., 20,000 employees) act similarly? After all, they have similar "bottom line" concerns, which are certainly as valid as those of the Fortune 100. And if they behave similarly, then how about small companies? And before you know it, aggregation is out the window, and that is A Very Bad Thing.

So we come up with Tony's solution to make NAT- or proxy-like additions to the IP architecture to account for this in order to localize the effects of renumbering to the perimeters, and thus reduce its pain. As Tony implied, it is much easier to do this at routers than the order of magnitude more numerous computers. However, such "fatter IP layer" approaches add latencies and impact performance, often in subtle ways. 

In my own mind, this topic is indirectly related to the goal of converging voice and video upon IP. I personally have concluded that QoS doesn't work as advertised as a convergence mechanism such that convergence can best occur in real life only via over-provisioning. I realize this is an unpopular conclusion, because it explicitly claims that the IP layer can only get so fat, after which it becomes ineffectual. I fear that multihoming will fatten the IP Layer in ways that we do not fully understand now. The degree of fatness of the IP Layer is something that must concern all of us... I am reminded of Steve Deering's many IETF plenary speeches back in the late 90s. This topic is a very weighty one.

If Tony Li is correct (and I explicitly trust his judgment), then the decision may be between keeping addresses constant and thus harming routing aggregation versus renumbering and thus fattening the IP Layer. Are there other alternatives? What do you think?

--Eric

-----Original Message-----
From: Brian E Carpenter [mailto:brc@zurich.ibm.com]
Sent: Friday, August 13, 2004 1:01 AM
To: Multi6
Subject: Re: Newbie Question about addressing impacts


Tony Li wrote:
> 
> On Aug 12, 2004, at 6:31 AM, Fleischman, Eric wrote:
> 
>> This is an interesting idea. However, if we embed proxy functions into 
>> border routers it would potentially add overhead (as well as latency) 
>> and make them harder to manage. Specifically, the number of border 
>> routers is likely to increase as network perimeters become more 
>> porous. Thus, this idea carries with it the need to ensure that these 
>> distributed routers can be configured with consistent policies.
>>
>> Simple is good in operations.
>>
> 
> 
> Well, then the other architectural alternative that I can see is to 
> embed NAT-like functionality in all of the hosts.
> 
> I find this scarier.

Chair hat off:

I repeat my comment from when I first saw Mike O'Dell's original 8+8
proposal: "It's architected NAT." I think anything that massages locators,
whether it's in the host stack or in a proxy, comes down to architected
NAT. Which means there is going to be state, so that the massage can be
reversed, so that the ULP always sees the same e2e identifier. It's a
design choice whether that state is in hosts, proxies, or both.

Actually, we're kidding ourselves if we don't admit that this is what
we are going to end up doing.

Chair hat on:

The design team has been asked to develop one specific approach
to this, namely the IP wedge layer approach, because that is where
the proposals and interest in the WG seem to be concentrated.

    Brian