Demystifying Performance of eBPF Network Applications#

Farbod Shahinfar; Sebastiano Miano (Politecnico di Milano); Aurojit Panda (NYU); Gianni Antichi (Politecnico di Milano & Queen Mary University of London)

• Why: eBPF is the load-bearing technology behind Cilium, service-mesh dataplanes, and a growing share of host networking. This paper asks the inverse of the usual question — not "how fast can eBPF go" but "where does eBPF actively hurt" — and the answer should change how you reason about which workloads belong on which side of the kernel boundary.
• What: A systematic benchmark study showing that many networked applications cannot benefit from eBPF at all, and that eBPF can create performance isolation violations between co-located applications and constrain deployment and workload choices. The paper also assesses whether in-flight kernel work (e.g., the ongoing JIT/verifier and scheduling improvements) fixes the identified limits, and proposes directions for the community (from abstract only).
• How: Multiple benchmarks under deliberately varied workloads with purpose-designed test cases isolating the failure modes — the methodology is the contribution as much as any single number.
• Where applicable: Broadly applicable to any team running eBPF in the host datapath (Kubernetes CNI dataplanes, host-LB offloads, socket-level accelerators). Treat the isolation-violation finding as a caveat for multi-tenant hosts specifically.