ReStore-in-Action: Adaptive Multi-Tier Storage Management via Intelligent Data Migration Policy
TL;DR
ReStore is presented, a lightweight, page-level Reinforcement Learning-based migration policy that adaptively optimizes data placement by considering both workload and storage device characteristics and achieves up to 6× lower runtime and 48× fewer migrations compared to state-of-the-art baselines.
Abstract
Modern data systems employ a diverse range of storage devices, from high-performance, expensive SSDs to capacity-oriented SSDs and cost-effective, high-capacity HDDs. To leverage these varying characteristics, storage devices are often organized into Multi-Tiered Storage (MTS) systems. In MTS, data placement must be dynamic to align with evolving workload patterns through a migration policy that continuously upgrades or downgrades pages between tiers. However, managing these movements is challenging, as traditional rule-based methods often lack the flexibility to handle complex workloads and fail to account for hardware-specific properties, such as SSD read-write asymmetry and internal parallelism. To address these limitations, we recently presented ReStore, a lightweight, page-level Reinforcement Learning-based migration policy that adaptively optimizes data placement by considering both workload and storage device characteristics. Experimental results using industry-grade benchmarks and reallife traces show that ReStore achieves up to 6× lower runtime and 48× fewer migrations compared to state-of-the-art baselines. In this demonstration, we present a web simulation of an MTS where we integrate ReStore along with five other baseline policies. The conference participants can configure storage tier properties, generate custom workloads, and observe ReStore's realtime behavior (data placement and migration between tiers) alongside the baseline approaches. The demonstration is available at: https://disc-projects.bu.edu/ReStore/research.html.