Orchestrating In-Band Data Plane Telemetry With Machine Learning

In-band network telemetry (INT) is an emerging network monitoring paradigm. By collecting low-level telemetry items in real time, INT can substantially enhance network-wide visibility - allowing, for example, timely detection problems such as micro-burst. Recent studies have focused on (i) developing INT mechanisms to increase network-wide visibility; and (ii) to design new monitoring solutions. However, little has been done to coordinate the process of collecting telemetry items in this new paradigm. This is particularly challenging because depending on which network telemetry items are collected, it might degrade network-wide visibility in terms of consistency/freshness. In this letter, we theoretically formalize the In-band Network Telemetry Orchestration Plan Problem and propose a machine learning based orchestration model. Results show that our approach outperforms state-of-the-art heuristics by up a factor of 8x with respect to the number of network anomalies identified, for instance.