High-Performance Peer-to-Peer Overlays, or, Heuristics Considered Harmful

In this talk, I will describe a new approach for building distributed systems with strong performance and resource usage guarantees. The critical insight behind this work is to formalize the core tradeoffs in distributed systems as a mathematical optimization problem. We can then achieve high performance in the presence of limited resources by minimizing a targeted performance function subject to constraints in a distributed fashion.

The talk will outline this new approach and describe how we recently applied it to build three peer-to-peer systems: CoDoNS, a replacement for DNS, CobWeb, an open-access content distribution network like Akamai, and Corona, an RSS-like system for disseminating Web micronews. All three systems have been deployed on PlanetLab, and either guarantee near-optimal lookup/update performance subject to bandwidth constraints or achieve a targeted level of lookup/update performance while minimizing bandwidth and storage costs. Overall, this approach represents a novel way to building large-scale distributed systems that contrasts with past systems based on ad hoc heuristics.
Gun Sirer is an assistant professor in the Computer Science Department at Cornell University. He works on self-organizing systems, which span operating systems, networking and distributed systems. Much of his research emphasizes building systems based on principled reasons for their correct functioning. His current projects involve peer-to-peer systems, systems support for ad hoc networks, and operating systems.