Impact of customizable orchestrator scheduling on machine learning efficiency in edge environments

This article explores the impact of custom scheduling strategies on the performance of machine learning workflows at the edge by using the case of Kubernetes scheduling. Optimizing machine learning (ML) workloads on resource-constrained edge devices has become a significant scientific challenge addressed by multiple studies. The severe limitations of edge systems in processing power, memory, and energy render conventional cloud-native schedulers ineffective, leading to poor resource utilization and degraded performance. While numerous advanced, data-driven solutions have been proposed for large-scale systems, their complexity and overhead are often impractical for edge deployments. In contrast, this work investigates a simpler, lightweight scheduling mechanism for CPU-based workloads that provides efficient and predictable performance without relying on historical data, making it well-suited for the unique requirements of the edge.Using a lightweight K3s cluster integrated with Kubeflow Pipelines, we investigate how varying binpacking functions influence resource allocation and training efficiency of a CNN model on the MNIST dataset. Our experiments demonstrate that tailored scheduling configurations can lead to noticeable improvements in training times and hardware utilization across different edge cluster sizes. The results offer actionable insights for optimizing AI workloads in resource-constrained edge environments.