Back from Netdev 0x19

Netdev 0x19Around mid-March, Bootlin engineers Köry Maincent and Romain Gantois had the chance to attend edition 0x19 of the Netdev conference, where developers from the Linux kernel networking community assemble for four days of technical presentations and discussions.

They were generally very much impressed by the quality and detail of the talks given during this conference, and they have selected two of the talks that they found the most interesting, which they summarized below. We expect the slides and videos to be published at some point in the future, but they don’t seem to be available at the moment.

Understanding and Improving Robustness of Networked Systems With Machine Learning

Talk by Brighten Godfrey

This keynote by Brian Godfrey, professor at UIUC, was focused on possible applications of Machine Learning for understanding and managing networked systems. Pr. Godfrey pointed out that while these systems were usually quite unpredictable once deployed, they presented the advantage of easily available performance metrics. According to him, this combination of properties was prime ground for Machine Learning techniques. To demonstrate this, he presented two examples where Machine Learning was applied to solve networked system management problems.

The first example was troubleshooting of incidents in networked systems with distributed applications. In such systems, especially when the number of different entities involved is quite high, inferring a root cause from a sudden local degradation of performances can be very difficult. Pr. Godfrey constructed a Machine Learning model called Murphy, which could be trained to learn a joint probability distribution on all performance metrics in the entity relationship graph. This model could then be used to generate “explanation graphs”, linking local changes in metrics to probable root causes in the graph of entities.

Talk abstract on the Netdev conference website

Challenges and limitations of debugging increasingly complex virtualized networks with Retis

Talk by Antoine Tenart, Paolo Valerio, Adrián Moreno

The session began by outlining the challenges of tracking packets through various layers, protocols, and drivers in Linux networking datapaths. This complexity complicates feature development and troubleshooting processes, making it increasingly difficult to diagnose and resolve networking issues effectively.

To address these challenges, the presenters introduced Retis, a tool that leverages eBPF’s flexibility and integrates with libpcap filtering. Retis is designed to monitor packets, collect contextual information from components like nftables, conntrack, and Open vSwitch, and facilitate in-depth analysis of the gathered data.

The tutorial delved into the architectural decisions behind Retis, shaped by the identified debugging challenges. Key features and implementation strategies of Retis were thoroughly examined, providing a comprehensive understanding of its capabilities.

Throughout the talk, multiple examples were used to demonstrate how Retis can be leveraged to gain insights into internal packet processing. These practical scenarios highlighted the tool’s effectiveness in troubleshooting and development environments.

The session concluded with a discussion on the current limitations of Retis and potential avenues for future enhancements, inviting collaborative input to refine the tool further.

Overall, the tutorial offered valuable perspectives on debugging complex networks and highlighted Retis as a promising solution in this domain. This tool will surely help us in debugging the networking stack, making network analysis and troubleshooting more efficient.

Talk abstract on the Netdev conference website

Conclusion

Overall, attending Netdev 0x19 was a highly enriching experience for our team. The depth and technical quality of the sessions reinforced the value of this conference for anyone involved in Linux kernel networking. We’re particularly excited to apply some of the insights gained in our future work. Many thanks to the NetDev Society for their efforts in organizing this great event! We’re already looking forward to the next edition.

Leave a Reply