[ppml] Policy Proposal 2005-1: Provider-independent IPv6

Vince Fuller vaf at cisco.com
Thu Apr 27 16:40:08 EDT 2006


Once again, I will submit myself to the definition of insanity by trying
again (and no doubt failing) to explain a little bit about Internet scaling.
In fairness, it is difficult to comprehend the issues and complexities of
large network scaling until one has worked first hand in a large network
environment.

> 2.	It is unlikely that the internet will see anywhere near the
> 	explosive growth of the 90s in the next 5-10 years. Even if
> 	did, we would still stay well short of 160,000 v6 routes which
> 	is well under most estimates I've heard for current hardware
> 	capability.  As such, there shouldn't be much of a problem
> 	for at least 10 years.

This "analysis" completely misses the point. IMHO, it is also simply wrong.

Growth trends clearly show that the fraction of Internet sites which are multi
homed and/or which engage in traffic engineering via more-specifics is on the
rise. Expect this trend to accelerate as the Internet becomes more mission-
critical for businesses of all sizes. A substantial fraction (20%? 50%?
more?) of the number of worldwide businesses would be a good guess at the
eventual number of multihomed sites. How many businesses today don't have
mission-critical phone service? There is good reason to believe that Internet
access will be just as important in the future, particularly if the vision
of the "converged network" comes to pass (not that I'm drinking that
kool-aide... but that is a separate conversation). Multi-homing will drive
the creation of enormous global routing state if the expedient "provider
independant space" approach is taken in favor of the more difficult task
of fixing the architecture (and no, I do not believe that shim6 will be
a viable alternative). Believe it or not, consumer Internet usage also
continues to grow at a super-linear rate, thanks to the emergence of
applications which appeal to mainstream users (for example, peer-to-peer
file sharing). As traditional telephony and other services are moved to
the packet network, this trend will accelerate.

As I probably mentioned in an earlier message, exponential and polynomial
growth curves often don't look very exciting during the shallow, early stage
of growth, when small numbers mask what will become an impressive trend.

> 3.	The large(ish) ISPs comprise the majority of the operational
> 	focus in the IETF, and, indeed have been a strong enough force
> 	there that they were able to get RFCs cranked out which
> 	attempted to preserve a completely provider-dependent
> 	addressing model for the v6 internet.  As such, faced with
> 	building a scalable routing system or waiting for the network
> 	to implode, I would hope that they will start working towards
> 	a more scalable solution, such as ID/LOC splits.

This statement belies a fundamental lack of understanding of the IETF and
the history of operator participation over the past 10 to 15 years. During
that period, the IETF has been almost completely devoid of operational focus,
partly due to low operator participation (see below for why some operators
feel alienated) and partly due to active bias against "evil greedy bastard"
ISPs. The output of the IETF, most notably in ipv6's lack of a scalable
routing and addressing architecture, reflects this. In that particular case,
those with operational experience were greeted with outright hostility; as
the political consensus-building process converged on ipv6-nee-SIP, some
operators did speak up and say "this isn't going to work... you are ignoring
the fundamental scaling problems in routing and addressing". They were told
"too bad" and actively discouraged from participating. Meanwhile the ipv6
brain trust said, to paraphrase a "talking Barbie" doll of the 1970s, that
"math [solving that problem] is hard, let's go shopping [design packet
headers] instead". 

> 5.	Vendor focus, in my experience, tends to be towards making
> 	the large(ish) ISPs happy and the majority of enterprises
> 	are a secondary consideration.  This makes sense when you
> 	consider that the average large(ish) ISP spends several
> 	million dollars per year with their router vendor(s) of
> 	choice, while the rest of the world is significantly less
> 	per enterprise (in most cases) spread over a much wider
> 	collection of sales representatives.  In most sales-oriented
> 	organizations (which as near as I can tell, all the hardware
> 	vendors are today), the sales rep with the largest dollar
> 	value tends to have the largest say in the feature priorities.

This statement couldn't be more wrong. In fact, I had to read it several
times to make sure that I had parsed it correctly.

I would be very interested in understanding your operational experience,
particularly as it pertains to large ISP networks. In my experience in 
working for regional, then national, then international ISP networks from
1988 until 2001, it was abundantly clear that equipment vendors viewed "that
Internet stuff" as a small, sometimes-profitable but low margin sidelight
to enterprise networking.  This was true when token ring source-route bridging
(remember that?) and IBM terminal emulation were implemented more quickly and
more correctly than sorely needed ISP features like OSPF, IS-IS, or BGP. It
remains true today.

Fact: high-end ISP equipment has large development costs, is intrinsically
expensive to build, sells in relatively low volume, and has specialized,
difficult, and/or costly-to-build features that appeal to a very small 
egment of the overall vendor customer base.

Fact: enterprise equipment, by contrast, has fewer, simpler features, 
far more customers, lower cost, and much higher margins.

Fact: ISPs may spend millions on equipment but enterprise customers spend
billions. These are American-English millions, with six trailing zeros, and
American-English billions, with nine trailing zeros (I understand that the
U.K.-English definitions may be different).

Fact: both sales and gross profit attached to enterprise sales dwarf ISP
sales by orders of magnitude.

Conclusion: enterprise customers drive vendor development, sales, and all
other priorities. This has been true since the years of Cisco, when HP,
Cisco's first big customer, bought orders of magnitude more equipment and
generated orders of magnitude more profit than all ISP sales combined.

It used to be said that the best motivator for development of ISP features
was a prediction that what ISPs are using today is what enterprise customers
will be using in 3 years. That meant that ISPs typically got those features
a year or two after they really needed them, but at least the eventually did
get them. It isn't clear whether the next generation of cutting-edge ISP
features will continue to "trickle down" to enterprise users; for ISP's sake,
one would fervently hope that they do.

This "conspiracy of the vendors and the big ISPs" has more in common with
analysis of the Zapruder tape and theories about the grassy knoll than it
does with business reality. It really is time to give it a rest already.

	--Vince
	(not in any way, shape, or form speaking for my current employer)



More information about the ARIN-PPML mailing list