LINT: Accuracy-adaptive and Lightweight In-band Network Telemetry

In-band Network Telemetry (INT) has recently emerged as a means of achieving per-packet near real-time visibility into the network. INT capable network devices can directly embed device internal state such as packet processing time, queue occupancy and link utilization information in each passing packet. INT is enabling new network monitoring applications and is currently being used in production for providing fine-grained feedback to congestion control mechanisms. The microscopic network visibility facilitated by INT comes at the expense of increased data plane overhead. INT piggybacks telemetry information on user data traffic and can significantly increase packet size. A direct consequence of increasing packet size for carrying telemetry data is a substantial drop in network goodput. This paper aims at striking a balance between reducing INT data plane overhead and the accuracy of network view constructed from telemetry data. To this end, we propose LINT, an accuracy-adaptive and Lightweight INT mechanism that can be implemented on commodity programmable devices. Our evaluation of LINT using real network traces on a fat tree topology demonstrates that LINT can reduce INT data plane overhead by ≈25% while ensuring more than 0.9 recall for monitoring queries trying to identify congested flows and switches in the network.