Evaluating Container Virtualization Frameworks for Edge-to-Cloud and Serverless Workflow Execution: Extending the DAGonStar Engine

Edge-to-Cloud workflows must run reproducibly across heterogeneous hardware, operating systems, and network conditions, yet most scientific workflow engines assume homogeneous, well-connected infrastructure. This paper reports on an extension of the DAGonStar workflow engine that adds container-based task execution through four virtualization backends — Docker, Apptainer, HashiCorp Nomad, and Kubernetes (via the lightweight k3s and k0s distributions) — together with a cross-cutting data-staging subsystem and an integrated Prometheus/Grafana monitoring stack. The extension is evaluated on a physical Fog–Edge–Cloud testbed built around a Raspberry Pi 5 edge node, a commodity PC acting as a Fog node, and a managed MongoDB Atlas cloud backend, using both synthetic file-generation/compression benchmarks and a complete IoT telemetry workflow driven by a DHT11 temperature/humidity sensor. Across repeated runs, Apptainer achieved the lowest execution time and energy consumption per unit of work (1.68 J/record) in single-node and end-to-end scenarios, Docker offered the best balance between throughput and portability in distributed Edge–Fog configurations, and Nomad, while the most stable in terms of storage growth, incurred the highest per-record energy cost (5.73 J/record) due to orchestration-agent overhead. Kubernetes, although fully supported at the software-integration level, could not be reliably operated on the resource-constrained edge device used in this study. This work was carried out as a Bachelor's thesis (Trabajo Fin de Grado) at Universidad Carlos III de Madrid (UC3M), condensed and restructured into technical-paper format. It is a self-archived preprint and has not undergone peer review. Code and experimental data: https://github.com/sebas80sebas/dagonstar-containers (DOI: 10.5281/zenodo.22801293)