Adaptive storage optimization for deep learning workloads in converged HPC-cloud environments: a reinforcement learning approach

Abstract

Storage systems remain a significant bottleneck in large-scale deep learning training. Most evaluation studies rely on synthetic benchmarks that do not capture the I/O heterogeneity present in real training pipelines. Automated configuration methods often treat parameter selection as stateless, overlooking the temporal structure of recurring training phases. These challenges are addressed using a converged HPC-cloud platform that integrates Lustre with S3-compatible object storage. Three workloads with distinct I/O characteristics were evaluated: ResNet-50 on ImageNet-1K, 3D U-Net on the Medical Segmentation Decathlon, and Whisper Large on LibriSpeech. Time-to-convergence was used as the primary metric. We introduced a deep reinforcement learning framework that combines Proximal Policy Optimization and Soft Actor-Critic with a Transformer-based state encoder to predict workload phase transitions rather than merely respond to them. The learned policies reduced checkpoint write latency by up to 74.6% and improved GPU utilization across all three workloads. The approach generalizes across heterogeneous I/O profiles without per-workload reconfiguration. The results indicate that temporally aware storage control can improve the efficiency of distributed training pipelines in heterogeneous HPC-cloud environments.