ONF Connect 2018

This talk describes an experiment that uses network telemetry collected via P4 INT and feed to a machine learning (ML) system to predict/detect network congestion state and use closed-loop control to dynamically modifiy the network to alleviate congestion.

The purpose of this experiment it so understand the capabilities that P4 INT enables and how using P4 INT as a telemetry data source compares to traditional network polling for telemetry.
As part of this experiment, the network load attributed to telemetry gathering will be studied to understand how / if P4 INT reduces the source (network and compute) for telemetry gathering. Additionally, the impact of fine grain v course grain telemetry on ML will be investigated.