Jennifer Rexford, Princeton University

Abstract

The proliferation of communication services that we depend on every day makes managing computer networks more important than ever. The increasing security, availability, and performance demands of these services suggest that these network-management problems must be solved in real time---and inside the network.  In this new era, network management requires a fundamentally new approach.  Instead of anomaly-detection algorithms that perform offline analysis of network traces, future networks need to make real-time, closed-loop decisions---to block unwanted traffic, to reroute traffic to avoid congestion, and more.  This talk explores how to bring together network measurement, analysis, and control, by leveraging recent advances in programmable network devices.  We present several example in-network "apps” (written in the P4 programming language) that detect and fix security and performance problems, while still processing packets at line rate in high-speed switches with limited memory. The resource constraints in modern programmable switches create new opportunities for research in streaming algorithms and compact data structures.

The three example “apps” presented in the talk are described in the following papers:

• Scalable load balancing (HULA)
• Heavy-hitter detection (HashPipe)
• Microburst diagnosis (Snappy)

Slides     Video