Distributed Probability Orchestrating for Probabilistic In-Band Network Telemetry

In-band telemetry (INT) is an emerging framework for network diagnosis, and probabilistic in-band telemetry (PINT) is its variant which consumes less bandwidth with a probabilistic sampling method. However, as we show in this paper, PINT causes non-uniform sampling probability at different switches within a topology, resulting in measurement accuracy not optimal. Therefore, it is meaningful to orchestrate probability in per-path and per-hop granularity. In this paper, we address this problem by formulating it as convex programming models and solve them. We also apply the solution to satisfy PINT's digest-overwritten fact. The method is implemented in software programmable switch with p4, a de facto standard for data plane programmability. Evaluation on a FatTree testbed shows that our method overperforms PINT with negligible extra bandwidth consumption.