RESOURCE‑AWARE MACHINE LEARNING FOR CLOUD–EDGE TASK ALLOCATION: A SMALL‑SCALE SYSTEM AND FEDERATED‑LEARNING IMPLICATIONS

Abstract

Heterogeneity in computation and communication across cloud and edge platforms presents a significant obstacle for task allocation. Heuristics that greedily assign tasks to the fastest worker can overload high‑capability nodes while leaving slower nodes idle. This paper presents Cloud‑Assisted Resource Allocation Using Machine Learning, a reproducible prototype that learns to allocate “cloudlets” to edge workers. The system comprises a synthetic data seeder, a cloud module that trains a decision‑tree classifier to predict the best worker for each cloudlet, and a master scheduler that uses the trained model to dispatch tasks subject to compute (MIPS) and network (bandwidth) constraints. We benchmark the machine‑learning (ML) scheduler against a greedy baseline and analyze per‑worker durations and makespan. On a representative run with three workers (W1=2.3 MIPS, W2=2.6 MIPS, W3=3.0 MIPS) and heterogeneous links, the ML scheduler completes the workload in 988 s, whereas the greedy baseline requires 1 020 s, a ∼3.2 % reduction in makespan and a substantial reduction in W3 overload. Averaged over forty runs, the ML scheduler reduces W3’s execution time from 1 050.5 s to 982.5 s, confirming consistent load balancing. Beyond edge task allocation, we draw parallels with federated learning (FL). The task → worker mapping resembles client → round selection in FL, and resource‑aware scheduling can mitigate stragglers and reduce time‑to‑accuracy. We discuss how the proposed prototype could be extended with systems such as Kubernetes Horizontal Pod Autoscaler[1] and KubeEdge[2], and we outline future work on integrating federated‑learning frameworks like Flower[3]