Published
Edge Computing Orchestration Overhead
First per-container decomposition of a Kubernetes distribution on real edge hardware — cgroup telemetry on a k0s worker attributing the orchestration tax across four system pods and its scaling under load (idle ~0.38% of node; control-plane churn 9.2× system CPU). ESOCC 2026, Springer LNCS
About
The first per-container decomposition of a Kubernetes distribution on real edge hardware. Using cgroups v2 telemetry on a k0s worker, it attributes orchestration overhead to the four worker-side system pods (kube-apiserver/konnectivity-agent, controller-manager, kube-proxy, CoreDNS) and tracks how the "orchestration tax" scales from idle through control-plane churn to a busy data plane. At idle one pod carries two-thirds of system CPU while the whole tax is ~0.38% of the node; control-plane churn multiplies system CPU 9.2×, and memory stays a near-fixed ~46–60 MiB band. Full title "Edge Computing Orchestration Overhead: Analysing Per-Container Decomposition of Kubernetes on Edge Using cgroup Telemetry", published in ESOCC 2026 (Springer LNCS) with Sergio Moreschini and David Hästbacka.