The WING Blog

The Web and Internetworking Group at BU/CS



Hourglass paper at SIGCOMM ’11

By John Byers

Easily the most widely discussed paper in hallway conversations at this year’s SIGCOMM was this modeling paper:

The Evolution of Layered Protocol Stacks Leads to an Hourglass-Shaped Architecture, by Saamer Akhshabi and Constantine Dovrolis (Georgia Institute of Technology).

For those of you who enjoy controversy, or who are prepping for this year’s DWE exam, I think reading this paper is worth your time.  Advocates for this paper thought that the topic is a fascinating one, and that trying to understand and model how an hourglass-shaped architecture arose is a worthwhile research line.  Detractors said (I’m paraphrasing using family-friendly language) that not only does the model fail to reflect or approximate reality, but the nature of an endeavor where you know the answer (hourglass) and contrive to build a model that produces the answer is bad science.

We decided to take this paper at the PC meeting in spite of its warts, since we knew it would lead to intense discussion and debate, and, as Jeff and I wrote in our opening remarks as PC co-chairs: “… we hope we took enough risky papers to get at least some of you upset.”  We succeeded!

3 Responses so far

Kudos to the PC for accepting this paper. I enjoyed reading the paper; it expands what we think about. I think people in the SIGCOMM networking community sometimes take too narrow a view of what constitutes valuable research. I see this paper as getting us to think more deeply about “what leads to a narrow waist”? After all, we’ve been talking about the narrow waist for 30 years but rarely go past superficial observations about it.

It think it misses the point to criticize this paper for not reflecting reality; but I agree with those who reject knowing what you are looking for and stopping when you find it.

OK, I have to read it first :) But just from the post, I don’t see where the “evolution” is if folks had started with only IP. Where was the competition??

The paper would make a lot more sense if the methodology is shown to explain other instances of the “hourglass” in layered abstractions — e.g., in software.

What explains the convergence of the software community to x86 narrow-waiste connecting PLs to ISAs?

Leave a comment