Abstract
Network Telemetry (NT) is a crucial component in today’s networks, as it
provides the network managers with important data about the status and
behavior of the network elements. NT data are then utilized to get
insights and rapidly take actions to improve the network performance or
avoid its degradation. Intuitively, the more data are collected, the
better for the network managers. However, the gathering and
transportation of excessive NT data might produce an adverse effect,
leading to a paradox: the data that are supposed to help actually damage
the network performance. This is the motivation to introduce a novel NT
framework that dynamically adjusts the rate in which the NT data should
be transmitted. In this work, we present an NT scheme that is
traffic-aware, meaning that the network elements collect and send NT
data based on the type of traffic that they forward. The evaluation
results of our Machine Learning-based mechanism show that it is possible
to reduce by over 75% the network bandwidth overhead that a
conventional NT scheme produces.