Making Decisions at Data Plane Speeds

Abstract

Feedback control loops to implement self-driving networks constitute data collection to sense the network, and control algorithms to make decisions driving the network. Highquality data is necessary for smart decisions. Yet, highquality data is hard to obtain from the network data plane, due to insufficient visibility and large data volumes stemming from high packet rates. This paper distills principles to collect high-quality data arising from our own research experience: (i) filter and aggregate data as close to the source as possible; (ii) identify broad families of statistics that are measurable with bounded inaccuracy; (iii) don't assume lowlevel data plane software is easy to instrument, but instead (iv) apportion software flexibility by the time scales of the computation; and (v) prefer in-band approaches where possible for timely and efficient reactivity. We call the community to act upon these principles to leverage emerging opportunities using safely-extensible network stacks.