isen.blog

Silicon Investor recently pointed to a paper [.pdf] by network researchers at AT&T, University of Nevada and Rensselaer Polytechnic that purports to address Network Neutrality by focussing on "the specific question" whether, "over-provisioning is an economically viable strategy due to the declining cost of capacity, instead of incurring the complexity and operational costs of running a differentiated services network."

SI quotes my former Bell Labs colleague and paper co-author KK Ramakrishnan as saying, "Clearly, an undifferentiated network in this context is less efficient and more expensive . . . We believe understanding the real impacts of the alternative strategies is important as the debate about network architecture unfolds."

I think the papers introductory assertion goes way beyond the scope of the study, and KK's quote, if indeed he is accurately quoted, goes way beyond the data presented.

There's nothing wrong with the study itself, as far as I can see. Looks clean. The research shows that a network that classless networks can require 60 to 100 percent more capacity than networks that use DiffServ to deliver traffic with equivalent delay and loss. Yawn.

But the paper does not address the costs of implementing DiffServ versus its substitute, i.e., more capacity. Does 60 to 100 percent more spare capacity equal 60 to 100 percent more cost? Not in my experience. My 20 Mbit FIOS connection costs me about 20% more than my 768 kbit DSL connection. Same for other measures of capacity; once you have the wire (or fiber or air) interface, power supply, protocol logic, right of way and customer interface, the difference between kilobits and gigabits is mostly a faster clock.

If you insist, two T1s probably cost twice what one costs. By the same token, it used to cost about one year's salary to sail across the Atlantic in 1750, and it *still* costs a year's salary to sail across, if that's how you insist on crossing the ocean. But that's not reality. T1 is as dead as the clipper ship.

The failure of the authors to extend the conclusions from capacity to raw costs of capacity is deliberately misleading, especially when the researchers invoked "economic viability" and "cost of capacity" in their introduction to the work.

Then there's "technological inflation," i.e., if you want to double your throughput without increasing your expenditures, all you've gotta do is wait a year (more or less). They ignore it.

But if you want to provision DiffServ, an engineeer hour this year has the same bang-per-buck as an engineer hour last year. Attendant human error probabilities also have a cost -- according to former BT CTO Peter Cochrane some 25% of network faults are caused by human error. Ignored.

Then there's the cost of denying service during peak traffic times to the poor innovator who can't afford the high-class service. What is the cost of inhibiting the next great idea? Ignored.

If you ask me, 100% over-provisioning is cheaper than ALL the other alternatives, and more beneficial to society too. The main advantage of engineering capacity as if it were scarce is that the telcos get to TREAT IT AS IF IT IS SCARCE by selling capacity to those who can afford it and denying it to those who can't.

Beware of geeks bearing diffs.

[Thanks to Frank Coluccio and Michael Bernstein for pointers to aspects of this work.]

UPDATE: Rudolf van der Berg had trouble posting a comment on this article in the usual way, so he emailed me to say:

You forgot another important element. Total traffic and peak traffic is growing at roughly 50-70% per year. So wowee that using QoS technologies might get you 50% efficiency. That's all gone by the end of the year. The trouble is that papers like these approach QoExperience in a static manner. The internet however is highly dynamic and growing, both endogenic (more traffic per user) and exogenic (more users).

Technorati Tags: AT&T, Economics, Framing, IntelligentNetwork, NetworkNeutrality, QOS, savetheinternet, Stupid Network