Proactive Data Placement in Heterogeneous Storage Systems via Predictive Multi-Objective Reinforcement Learning
Modern data-intensive applications demand efficient orchestration across heterogeneous storage tiers, ranging from high-performance DRAM to cost-effective cloud storage. Existing tiered storage systems predominantly employ reactive policies that respond to observed access patterns, leading to subopt...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11072103/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Modern data-intensive applications demand efficient orchestration across heterogeneous storage tiers, ranging from high-performance DRAM to cost-effective cloud storage. Existing tiered storage systems predominantly employ reactive policies that respond to observed access patterns, leading to suboptimal performance under dynamic workloads and failing to address multi-objective optimization requirements. We propose a novel proactive data placement framework that integrates predictive deep learning with multi-objective reinforcement learning to anticipate future data access patterns and optimize placement decisions across storage hierarchies. Our method employs Long Short-Term Memory networks and Transformer architectures to model complex temporal dependencies in I/O traces, generating predictive access probability distributions for data blocks. A deep reinforcement learning agent subsequently leverages these predictions, along with application-specific metadata hints, to make proactive placement decisions that simultaneously optimize latency, throughput, and cost objectives. The system incorporates a sophisticated reward mechanism that balances performance gains against migration overhead, while employing prioritized experience replay and adaptive learning rates to handle non-stationary workload characteristics. Through comprehensive evaluation using both synthetic and real-world traces from deep learning training workloads, our method demonstrates substantial improvements over state-of-the-art algorithms: achieving up to 45.1% reduction in average I/O latency, 32.5% improvement in throughput for critical applications, and 28.8% reduction in storage costs. The framework’s ability to proactively adapt to evolving access patterns while maintaining computational efficiency makes it particularly suitable for large-scale machine learning and scientific computing environments where data placement critically impacts overall system performance. |
|---|---|
| ISSN: | 2169-3536 |