A Distributed Open Software Router: Analysis and Architectures

The first generation routers were essentially general purpose computers performing packet processing in software. They executed routing protocols on UNIX like operating systems and the same processor was shared for all the different routing functions. With the advent of faster and more complex networks routers evolved from open and easily programmable computers to closed and specialised boxes. It has been shown that a general purpose processor can indeed provide very high performance in a router [1]. Motivated by this and other related studies, we focus on providing high performance routing function using open hardware and open source software. Specifically, we concentrate on distributed implementations using the most widely available open platform for this purpose—the general purpose processor based PC running open software. Several bodies of work have improved the performance of a single PC based software routers running Linux. Much of this work has focused on improving the software of the operating system, network stack, routing protocols and device driver optimisations. The central assumption in these works is that the CPU is the bottleneck for the entire system. Two proposals for loosely coupled distributed open software routers are also based on this assumption in their designs. A PC has three different subsystems, i.e., CPU, memory and I/O, that interact with each other during packet processing. In the first part of our work, we design and execute tests that stress each of these three subsystems. A principal observation from our work is that the complex interplay between the CPU, the memory and the I/O subsystems give rise to different bottlenecks under different operating regimes. We find that the I/O is the main bottleneck. We also find that the memory subsystem becomes a bottleneck for large routing tables. In the second part of our work, we build on these observations to propose several easy-to-implement and easy-to-configure distributed architectures. We first evaluate two loosely coupled distributed systems and understand their performance and architectural issues and then analyse the issues