Pricing and Internet Architecture

Frisky070802 writes "The Politech list recently posted a pointer to a new paper (pdf) by UMN prof Andrew Odlyzko, which compares the telecom industry to the historical transportation industry (railroad, bridges, and such). One quote, from the conclusion, is particularly interesting: '... the networking industry [has] devoted inordinate efforts to technologies such as ATM and QoS, even though there was abundant evidence these were not going to succeed. One can go further and say that essentially all the major networking initiatives of the last decade, such as ATM, QoS, RSVP, multicasting, congestion pricing, active networks, and 3G, have turned out to be duds. Furthermore, they all failed not because the technical solutions that were developed were inadequate, but because they were not what users wanted.'"

Comments

HOWTO: Write a stupid paper

[jd]

• First, pick a technology that has never been seriously deployed. It greatly helps if none of the readers have any practical, real-world experiences with the subject.
• Second, always compare with an only marginally related industry or discipline. There has to be enough of a connection to convince the readership, but not enough of one to disprove your preconceived notions.
• Third, Always tell the audience that they have The Right Thing. It makes them happy. It also makes the usual sponsors of such work (the ones who run the status quo) very happy, too. Happy enough to pay you, for example.

Seriously, multicaasting is enabled on most of the majopr backbones, but none of the major ISPs supply it - even to broadband customers - at any price. UUnet is one of the few that does. Their links aren't cheap, and from all accounts it can be very hard to get multicasting enabled, simply because a good number of their front-line support people don't know anything about it.

QoS is likewise serioulsy hindred. Oh, it's used in the field. The transatlantic link between the UK and the US has CBQ (Class Based Queues) enabled to maximise the throughput of important traffic, simply because there's so much.

Britain's JANET network has a highly extensive network of web caches. The theory being that one of the biggest loads on the transatlantic link is web traffic, and that the same site is often accessed repeatedly (eg: for University coursework), so that the most efficient solution is to cache everything.

While not strictly "QoS", caching can reduce access times for a web page at peak time from maybe an hour to down to 15 seconds, whilst also massively reducing the load on the network.

RSVP is a different case. That is known to not scale well over very large, complex multicast networks. (Too much overhead.) However, it is great for local networks, and I'm sure that it will gradually filter its way into Universities and mid-sized corporations, where videoconferencing is useful but bandwidth issues make it impossible to do without some QoS.

ATM is used by many xDSL companies, as it is a very efficient way to run a fixed-point to a fixed-point. To say it's not used is absurd and shows a degree of ignorance. It's also very popular in Europe, where people perhaps put a little more investment into infrastructure.

Quick note: I'm a little irritated by hearing some American politician label maglev trains as "sexy science fiction" and "stupid". To me, it's part of a worrying trend I'm seeing in all too much of the US, where there is an apparent phobia of making any actual progress in anything. To me, progress is the certain bit. What happens to those who reject it - that's not so certain.

How does this fit in? There's only so much bandwidth. Sure, Lucent is up to 3 Terabits per second, but with collapsing R&D funds and Lucent in enough of a financial mess, don't expect either a rollout, or a refinement, any time soon.

That's the absolute upper cap. The real limit is much smaller. Backbone connections are probably not much more than four or five hundred gigabits per second. (That is to say, about the capacity of two or three hundred well-made Pentium IV-based PCs.)

A relatively small Beowulf cluster could totally saturate a decent chunk of the Internet backbone. Most cluster-based computers, such as the Origin 3000 or the Altix 3000, with sufficient network links, could easiy max out the capacity of any part of the Internet, without much effort.

Why isn't the technology used? Because the customer doesn't want it? The customer has never been offered it!! Very, very few customers even know about it!! And ISPs, in particular, are keen to keep it that way. There is much more money to be made from serving people badly, because the customer'll keep paying for improvements and/or support. The ISPs can gouge the more foolish for years