Online learning on the programmable dataplane

This thesis makes the case for managing computer networks with datadriven methods automated statistical inference and control based on measurement data and runtime observations—and argues for their tight integration with programmable dataplane hardware to make management decisions faster and from more precise data. Optimisation, defence, and measurement of networked infrastructure are each challenging tasks in their own right, which are currently dominated by the use of hand-crafted heuristic methods. These become harder to reason about and deploy as networks scale in rates and number of forwarding elements, but their design requires expert knowledge and care around unexpected protocol interactions. This makes tailored, per-deployment or -workload solutions infeasible to develop. Recent advances in machine learning offer capable function approximation and closed-loop control which suit many of these tasks. New, programmable dataplane hardware enables more agility in the network— runtime reprogrammability, precise traffic measurement, and low latency on-path processing. The synthesis of these two developments allows complex decisions to be made on previously unusable state, and made quicker by offloading inference to the network. To justify this argument, I advance the state of the art in data-driven defence of networks, novel dataplane-friendly online reinforcement learning algorithms, and in-network data reduction to allow classification of switchscale data. Each requires co-design aware of the network, and of the failure modes of systems and carried traffic. To make online learning possible in the dataplane, I use fixed-point arithmetic and modify classical (non-neural) approaches to take advantage of the SmartNIC compute model and make use of rich device local state. I show that data-driven solutions still require great care to correctly design, but with the right domain expertise they can improve on pathological cases in DDoS defence, such as protecting legitimate UDP traffic. In-network aggregation to histograms is shown to enable accurate classification from fine temporal effects, and allows hosts to scale such classification to far larger flow counts and traffic volume. Moving reinforcement learning to the dataplane is shown to offer substantial benefits to stateaction latency and online learning throughput versus host machines; allowing policies to react faster to fine-grained network events. The dataplane environment is key in making reactive online learning feasible—to port further algorithms and learnt functions, I collate and analyse the strengths of current and future hardware designs, as well as individual algorithms