Classical analysis of when and how to apply basic engineering design techniques. Read this one on your own, skim it first, go for depth later.
A very readable and still relevant discussion of a fundamental design decision in the TCP/IP protocol design. A classic (although security concerns have evolved a bit since then).
A brief update on how the End-to-End argument impacts more modern system design.
Active Networks is an approach where the nodes in the network do some processing of traffic for efficiency or to provide services that are difficult to implement in an end-to-end fashion. This paper challenges blindly sticking to an end-to-end approach and provides some analysis.
A ground breaking paper which empirically refuted the base assumptions made in Poisson Process Models of internet traffic flow (that is that the interarrival and service time distribution assumptions do not hold in large networks).
A ground breaking paper which empirically refuted the base assumptions made in Poisson Process Models, and confirmed the presence of self-similar (long range time dependent) behavior of internet traffic flow. This takes a more mathematical approach.
A more informal introduction to the topic, it might be good to read this first as an entry point to get a qualitative feeling for what the problems are.