How to Use In-Band Network Telemetry Wisely: Network-Wise Orchestration of Sel-INT

As a promising network monitoring technique, in-band network telemetry (INT) helps to visualize networks in a fine-grained and real-time manner. Meanwhile, to address the overheads of INT, people have proposed a few selective INT (Sel-INT) approaches that only select a portion of packets in each flow to insert INT fields and distribute different types of INT fields over the selected packets. In this paper, we study how to use Sel-INT wisely in a network such that the tradeoff between monitoring accuracy/coverage and INT overheads can be balanced well. Specifically, we try to orchestrate the Sel-INT schemes of flows in both network- and flow-levels. For the network-level optimization, we model it as an INT planning problem in which the Sel-INT schemes of flows should be determined to maximize the information gain of INT as well as minimize the bandwidth overheads of INT. We formulate an integer linear programming (ILP) model to tackle the problem, prove its
$\mathcal {N} \mathcal {P}$
-hardness, and leverage Lagrangian relaxation to design a polynomial-time approximation algorithm for it. The flow-level optimization considers a dynamic network environment, and we propose to combine deep learning (DL) based traffic prediction with Sel-INT, such that the Sel-INT scheme of each individual flow can be updated timely and adaptively. We implement the proposal in a small but real network testbed and experimentally demonstrate self-adaptive orchestration of Sel-INT with it.