2026-10-05T09:34:26Z https://drops.dagstuhl.de/oai

oai:drops-oai.dagstuhl.de:25423 2026-03-19T12:46:44Z ddc:004 open_access

SEKHMET: Hash-Chained Perception Contracts for Heterogeneous Real-Time Edge Clusters El-Hadedy, Mohamed edge clusters K3s Kubernetes real-time perception scheduling integrity contracts hash chaining Hailo-8L Real-time perception pipelines on edge clusters are often scheduled as ordinary latency-sensitive pods, even when safety depends on sustained throughput and stable model outputs. This paper presents SEKHMET (Scheduling Edge Kubernetes with Hash-chained Monitoring of End-to-end Telemetry), a perception-aware orchestration layer for lightweight Kubernetes (K3s) clusters that exports window-level perception status as a control-plane signal. SEKHMET evaluates a perception-integrity contract (PIC) once per fixed-duration window and commits each window outcome into a hash-chained perception root that is published to an otherwise unmodified K3s control plane. The prototype uses a Raspberry Pi 5 perception-root node with a Hailo-8L accelerator, USB camera, and GPS receiver running a YOLOv8s detector, while up to five additional nodes generate elastic interference via swarm-stress. Under contract-unaware scheduling with multi-node interference, the end-to-end perception loop delivers ∼0.8-2.2 FPS and violates the PIC timing requirement in most of 214 windows, despite apparently healthy CPU and memory metrics. Under the same and heavier interference, SEKHMET sustains 27-30 FPS with contract_ok = True across 400 protected windows while publishing one 96-byte record per T=5s window (19.2 B/s control-plane payload). These results show that making perception requirements control-plane-visible can turn fragile best-effort perception into a protected cluster-level resource on commodity edge hardware. Schloss Dagstuhl – Leibniz-Zentrum für Informatik Mohamed El-Hadedy 2026 Is Part Of OASIcs, Volume 140, 7th Workshop on Next Generation Real-Time Embedded Systems (NG-RES 2026) InProceedings Text doc-type:ResearchArticle publishedVersion application/pdf doi:10.4230/OASIcs.NG-RES.2026.5 urn:nbn:de:0030-drops-254239 https://drops.dagstuhl.de/entities/document/10.4230/OASIcs.NG-RES.2026.5 eng https://creativecommons.org/licenses/by/4.0/legalcode