Efficient Scaling on GPU for Federated Learning in Kubernetes: A Reinforcement Learning Approach

Distributed learning enables efficient training with large-scale data, allowing processing to occur in multiple locations without centralization. However, the "straggler problem", referring to delayed update due to performance heterogeneity, hinders efficient learning. GPUs are crucial for enhancing learning speed by efficiently processing large-scale data through parallel computing capabilities. Efficient GPU utilization, considering resource costs, is essential for reducing training time and costs. Meanwhile, Kubernetes offers scalability, management ease, and resource provisioning for efficient operations. In this paper, we propose a Kubernetes-based vertical scaling scheme to address FL's straggler problem. For GPU scaling, we formulate an optimization problem that considers both resource cost and learning speed. Then, we propose a DRL-based approach to address this problem. We also leverage various Kubernetes features as well as CUDA Multi-Process Service (MPS) for execution of vertical scaling. We validate the proposed scheme's performance through various evaluations on a real testbed.