Thesis Defense - Killian Castillon du Perron (Université Côte d'Azur, Laboratoire i3S) - 21/11/2025

Value of the Information & Strength of the Argument

The value of the information is high, as it provides original empirical data on AF_XDP latency and CNI performance, filling a gap in existing literature. The argumentation is solid, with a clear methodology: automated benchmarking, clustering analysis, and a novel topology generator. The speaker systematically compares results and draws conclusions based on data. The optimization for Cilium is well-motivated and evaluated with concrete metrics. The presentation is logical and builds on previous contributions.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is strong, with a peer-reviewed publication (ICC 2024) for the AF_XDP study. The speaker references relevant prior work, such as the microservice topology study, and uses standard tools like the AF_XDP sample application. The title accurately reflects the content. The defense is well-structured and transparent about limitations, such as the complexity of AF_XDP and the need for fallback for security policies.

The title accurately reflects the content, which covers acceleration of container networks with eBPF, including performance characterization, evaluation, and optimization.

8/10

The defense presents original research with a rigorous methodology, including extensive benchmarking (over 100 configurations for AF_XDP, 170 topologies for CNI comparison) and a peer-reviewed publication (ICC 2024). The speaker is a PhD candidate, and the work is supervised by established researchers. The presentation is detailed and transparent about methods and limitations.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and context of the thesis on container network acceleration with eBPF.
• 02:00Explanation of microservices and container networking challenges, including latency issues.
• 05:00Introduction to eBPF and XDP, and the AF_XDP mechanism for bypassing the kernel network stack.
• 10:00First contribution: AF_XDP latency study, methodology, and results showing 6.5 µs minimum latency.
• 20:00Second contribution: CNI comparison (Antrea, Cilium, Flannel) using a microservice topology generator.
• 30:00Third contribution: XDP acceleration for VXLAN overlay in Cilium, achieving up to 28.8% latency reduction.
• 40:00Conclusion and future work, emphasizing the importance of both infrastructure and application architecture.

• EUR DS4H funded theses — Mentioned in the video description as a link to learn more about theses funded by EUR DS4H, which funded this thesis.

• Understanding Delays in AF_XDP-based Applications (ICC 2024) — The speaker mentions this publication as the basis for the first contribution, but no URL is provided.