Reliable and Distributed Network Monitoring via In-band Network Telemetry

TL;DR

This paper introduces GPINT, a reliable distributed network monitoring framework using In-Band Network Telemetry, which improves data collection reliability and efficiency over multi-hop paths in programmable networks.

Contribution

It extends previous INT work by integrating a shared queue ring for reliability, optimizing INT paths, and implementing the scheme in P4 for enhanced network telemetry robustness.

Findings

GPINT outperforms traditional methods in latency and overhead.

Reliability features significantly reduce data loss during congestion.

Emulation shows robustness under high packet loss conditions.

Abstract

Traditional network monitoring solutions usually lack of scalability due to their centralized nature collecting heartbeats from all network components via a single controller. As a solution, In-Band Network Telemetry (INT) framework has been recently proposed to collect network telemetry information more autonomously and distributedly by employing programmable switches. However, it imposes further challenges to (i) find suitable INT paths to optimize the control overhead and information freshness and (ii) ensure reliable delivery of control information over multi-hop INT paths. In this work, we propose a monitoring scheme, reliable Graph Partitioned INT (GPINT), by extending our previous work and integrating shared queue ring (SQR) as a reliability feature against potential failures in network telemetry collection due to network congestion and link degradation that may cause loss of the visibility of the network. We implement our proposal in a recent data plane programming language P4, and compare it with traditional Simple Network Management Protocol (SNMP) and also another state-of-the-art study employing Euler's method for INT path generation. Our analysis first shows the importance of having a data recovery mechanism against packet losses under different network conditions. Then, our emulation results indicate that GPINT with reliability extension performs much better than its opponent in terms of telemetry collection latency and overhead monitoring scheme even under a high amount of packet losses.