The Trade-Offs of Intelligent Orchestration in Kubernetes: A Multi-Infrastructure Case Study with the CODECO Toolkit
Authors:
- Georgios KoukisGKGeorgios Koukis
- Ioannis DermentzisIDIoannis Dermentzis
- Vassilis TsaoussidisVTVassilis Tsaoussidis
- Jan LenkeJLJan Lenke
- Fabian WölkFWFabian Wölk
- Daniel UcedaDUDaniel Uceda
- Guillermo SánchezGSGuillermo Sánchez
- Miguel Angel Puentes VallejoMVMiguel Angel Puentes Vallejo
- Javier SerranoJSJavier Serrano
- Panagiotis KaramolegkosPanagiotis Karamolegkos
- Rute C. SofiaRute C. Sofia
Open Access
- 5 November 2026
- journal article
- research article
- Published by MDPI AG in Applied Sciences
- Vol. 16 (21) , 10988
Abstract
IoT monitoring systems, from traffic and pedestrian sensing in smart cities to building energy metering and industrial robot telemetry, increasingly run as containerised microservices on Kubernetes (K8s) clusters that span constrained edge nodes and cloud servers. Cognitive orchestration layers extend K8s with network-, data- and energy-aware placement, running on the same nodes as the monitoring workloads they manage, with their own footprint rarely reported. This paper proposes a component-level cost-attribution methodology and an automation-footprint trade-off analysis that relates the manual interventions removed to deployment latency, CPU, memory and estimated power consumed by each of its functional roles. We apply it to Cognitive Decentralised Edge–Cloud Orchestration (CODECO), an open-source cognitive orchestration framework, across six infrastructures, including three operational IoT deployments, ARM and x86 hardware, and vanilla K8s, K3s and KinD. The automated CODECO and CODEF workflow reduces the number of manual interventions from 72 to 9 (87.5%) under our action-count definition, with a total of 4.57 GB memory overhead, the most significant observed trade-off. Decomposing these costs by architectural layer reveals that within the tested cluster sizes, the installation time of the scheduling component shows no measurable growth and that of the learning component grows slowly, while cross-layer network management contributes the largest share of installation time and idle CPU consumption. Results show that orchestration costs arise in different architectural layers, indicating they can be placed and optimised independently without sacrificing automation benefits.Keywords
This publication has 40 references indexed in Scilit:
- An autonomous energy-aware resource scheduling mechanism in serverless edge computingArtificial Intelligence Review, 2026
- A Framework for Cognitive, Decentralized Container OrchestrationIEEE Access, 2024
- Impact of power consumption in containerized clouds: A comprehensive analysis of open-source power measurement toolsComputer Networks, 2024
- An Empirical Evaluation of the Energy and Performance Overhead of Monitoring Tools on Docker-Based Systems2023
- Dynamic, Context-Aware Cross-Layer Orchestration of Containerized ApplicationsIEEE Access, 2023
- Shoggoth: Towards Efficient Edge-Cloud Collaborative Real-Time Video Inference via Adaptive Online Learning2023
- Kepler: A Framework to Calculate the Energy Consumption of Containerized Applications2023
- COGNIT: Challenges and Vision for a Serverless and Multi-Provider Cognitive Cloud-Edge Continuum2023
- Kubernetes distributions for the edge: serverless performance evaluationThe Journal of Supercomputing, 2022
- Edge Computing in Industrial Internet of Things: Architecture, Advances and ChallengesIEEE Communications Surveys & Tutorials, 2020