Tomorrow's networks are becoming more intelligent in several key ways. The first is that switch and NIC hardware have become considerably more programmable—allowing for greater network-application cooperation, such as in-network services, complex telemetry and data processing at scale, and more optimal transports. However, all aspects of the network, from the endpoints to the fabric itself, are still governed by heuristic algorithms and tuning variables that require in-depth expertise to develop and deploy. The second way that networks have become more intelligent is via recent advances in machine and reinforcement learning, which present promising methods to automatically infer policies for scheduling, control, and classification using past data, simulations, and live networks.
At the intersection of these two kinds of intelligence—programmable dataplanes and data-driven networking—lie new ways to better run future networks. We have been examining this from several angles, in how data-driven techniques might help the network and in how the network fabric can enable new ML-based analysis:
Revisiting the Classics: Online RL in the Programmable Dataplane – Kyle A. Simpson, Dimitrios P. Pezaros. (NOMS '22)
Poster: Online RL in the Programmable Dataplane with OPaL – Kyle A. Simpson, Dimitrios P. Pezaros. (CoNEXT ‘21)
Seiðr: Dataplane Assisted Flow Classification Using ML – Kyle A. Simpson, Richard Cziva, Dimitrios P. Pezaros. (Globecom '20)
Per-Host DDoS Mitigation by Direct-Control Reinforcement Learning – Kyle A. Simpson, Simon Rogers, Dimitrios P. Pezaros. (IEEE TNSM)