The Internet is huge, complex, and rapidly evolving.  Understanding how today's Internet-scale systems work is challenging, but crucial when designing the networks and applications of tomorrow.  In this

talk, I will describe how I have used a combination of measurement, modeling, and analysis to understand two Internet-scale systems: (1) peer-to-peer (P2P) file-sharing systems and their workloads, and (2) indirection routing systems that recover from Internet path failures.

In part because of the rise in popularity of P2P systems, multimedia workloads have become the dominant source of Internet traffic.  Our measurements show that multimedia workloads are substantially

different from traditional Web workloads.  Based on an analysis of a 6-month long trace of the Kazaa P2P system, I will propose a new model for multimedia workloads and will use it to explain how a few, simple, fundamental factors drive them.

In the second part of my talk, I will focus on understanding Internet path failures and indirection-based recovery schemes.  I will first characterize the frequency and location of Internet path failures that occur in practice.  Using insights drawn from our measurements, I will show how a simple, stateless, and scalable scheme called "one-hop source routing" achieves close to the maximum possible recovery

attainable by any indirection routing scheme.  I will also relate experiences we gained from implementing and deploying one-hop source routing on PlanetLab.