Comparison of distributed Machine Learning frameworks in a fog environment: Conceptual and Performance analysis

Abstract

The growing demand for real-time, latency-sensitive, and privacy-preserving analytics in IoT has brought fog computing as an alternative to cloud-based processing. However, training machine learning and deep learning (ML/DL) models in fog environments remains challenging due to limited computational resources. Despite the availability of numerous distributed ML frameworks, there is a lack of a comprehensive evaluation tailored to fog devices. This study conducts a comparative analysis of distributed ML frameworks for neural network training on resource-constrained fog nodes, using Raspberry Pi (RPi) devices. We started with Actor programming model-based frameworks and the study extended to general purpose distributed frameworks suitable for fog computing devices. We evaluate four actor-model-based frameworks (Akkordeon, DistBelief with Akka, Aktorain, and CANTO) along with general-purpose distributed frameworks (KubeRay, TensorFlow MultiWorkerMirroredStrategy (MWMS), Dask Distributed and Spark with Elephas). The frameworks are compared across key metrics including training time, accuracy, and resource utilization on diverse datasets. Our results highlight performance trade-offs: KubeRay offers a balance between efficiency and performance, Dask and MWMS achieve higher accuracy with increased latency, while Spark with Elephas excels in speed but struggles with accuracy. Although CANTO is optimized for fog-based training, it faces challenges with complex datasets. Overall, KubeRay emerges as the most practical choice for fog-based ML training because of its additional support for scalability and fault tolerance. This work bridges a critical research gap by providing experimental insights into the feasibility and performance of distributed ML frameworks in fog computing environments.