Nalanda

Citation: S. Shenker, Fundamental design issues for the future Internet, IEEE J. Selected Areas Comm., 13 (1995)

I found this paper really useful for two reasons. First, Shenker provides a view of what the future Internet might look like from 1995, anticipating widespread use of real-time applications. Second, and more importantly perhaps, he gives us a set of abstractions by which to judge network architectures based on different the demands imposed by elastic and real-time applications.

The applications in common use on the Internet at the time were “elastic”, in the sense that they were tolerant of delays in transmission, with design decisions that were motivated by the packet-switched nature of the Internet. It is, however, much harder to meet the demands of real-time applications, such as voice and video streams, using a packet-switched network, as they are much less tolerant to delay and loss.

Shenker demonstrates that a unified network infrastructure would provide the highest utility. To allow different kinds of applications to coexist, he considers overprovisioning, and modifications of the basic Internet architecture to provide a variety of service models beyond just best-effort service, which may also depend on admission control. The general intuition at work here is that aligning services with application requirements results in better overall network utilization.

Two models for assigning service to applications are presented: an implicit model where the network attempts to guess an application’s requirements, and an explicit model, where an application can declare its requirements to the network. The implicit model can be introduced with no changes to current service models, but may not be able to service applications that it does not know about, or variations in an application. The explicit model, on the other hand, requires a new service model (and therefore changes to the Internet architecture), stable service offerings, and also raises interesting questions of incentives: how can applications be incentivized to seek the levels of service they actually require, rather than always seeking the highest available level? Shenker suggests that they only way to address this concern is through variable pricing for different classes of service, though raising the concern that such variable pricing might change the nature of the gift economy prevalent on the Internet.

The most interesting section of this paper, for me, was the discussion of utility curves for different classes of applications: elastic, hard real-time, delay-adaptive and rate-adaptive. The shifts between concave and convex shapes, and the inflection points at which these occur, provide a compelling framework for thinking about the design of different classes of service in a unified network.