Fundamental Design Issues for the Future Internet

Looking at the future from 1995, about the time the Internet was being commercialized, this paper examines design issues going forward from services that are latency-tolerant ("elastic" in their designations) to services like video and audio, which require a better "class" of service. The paper advocates changing the network to accommodate different service types. Building a framework to evaluate different implementations, the paper introduces a measure that maximizes the "utility" of all flows--- with some flows requiring real-time service and others with less strict requirements.

A result of this paper is that sharing between flows that need fast service and elastic flows actually is more efficient than having separate networks for the two types of service. In addition, it is unrealistic to have the network "detect" the required type of service, so the network has to expose some kind of interface for the applications and users to request the types of service.

Two major problems with having different kinds of service is 1) how do you get people to settle for lower-levels of service and 2) how do you get ethernet/ATM/etc to support different classes of service. The first problem is fundamental in that one must incentivize things such that higher levels of service cost more--- the author rightly points out that this is going to require a huge culture change. In fact, the culture change is so large that I'm not sure it could be brought about, except if it was through services people already pay for (cable tv, for example). Asking for more money for services users are used to having for free is pretty insurmountable in the context of the internet.

Moreover, as Shenkar points out over and over, his analysis is based on bandwidth continuing to be expensive. Many of these concerns go away if bandwidth can be increased by a huge amount; however, I think the desire for low-latency class of service still remains today, but for applications such as gaming, the issue is not high bandwidth, but just having low latency. Thus there is a spectrum of latency and bandwidth requirements that would benefit from having different classes of service.

But while the Internet remains "good enough," this won't happen, especially given the invasiveness of the changes required. Even Ethernet wouldn't work. I don't have much critical to say about the paper, as it is more setting up a framework, one that I believe is reasonable given the concerns at the time.