Understanding and Surviving Internet Path Failures

Опубликовано 6 сентября 2016, 5:24

The Internet has experienced an exponential growth in recent years, so has its complexity, which can lead to more network-layer instabilities. What we urgently need today is a better understanding of these anomalies and more robust network protocols that perform well during failures. In this talk, I will describe two projects related to improving the robustness and efficiency of the Internet. The first is to build systems and tools that help understand the Internet's behavior during failures. We propose new techniques to observe wide-area network behavior by monitoring wide-area services, such as peer-to-peer systems and content distribution networks. Our system PlanetSeer, a novel diagnostic tool, passively monitors network communications for anomalies, and engages widely-distributed probing machinery when suspicious events occur. PlanetSeer detects anomalies at a rate that is 10 to 100 times higher than previous work, while incurring little probing cost. The second is to design end-to-end network protocols that are resilient to misbehavior. We design a novel transport layer protocol that makes use of multiple paths in parallel. We propose algorithms to support per-path congestion control and loss recovery, identify disjoint network paths, and suppress paths with shared congestion. Our system is able to achieve both higher throughput and shorter recovery time during path failures than standard TCP.