Carbon-aware Storage Tiering Framework

Appears in:

aism 26(1):

Authors:

Houman Kosarirad, Seyed Hossein Taheri, Nedasadat Taheri

Published:

8th Oct 2026

Abstract:

The rapid growth of cloud services and serverless applications has significantly increased the demand for efficient storage management in modern data centers. Tiered storage architectures, which combine high- performance solid-state drives (SSDs) with high-capacity hard disk drives (HDDs), are commonly used to balance performance and cost. However, accurately identifying which data objects should reside in fast storage tiers remains a challenging task due to dynamic and skewed access patterns in real-world workloads. In response, this paper introduces a predictive and carbon-aware storage management framework designed to improve data placement decisions in tiered storage systems. Our approach analyzes blob access traces and employs workload prediction to estimate future data hotness, enabling proactive migration of frequently accessed objects to high-performance storage tiers. By integrating prediction with migration cost awareness and time-varying grid carbon-intensity signals, the framework defers eligible migrations during carbon- intensive periods and prioritizes them during lower-carbon periods. Evaluation on an activity-focused subset of the Azure Functions Blob Access Trace, together with larger-scale experiments using up to 1,000 blobs, shows that the carbon-aware policy reduces operational carbon emissions by 27.76% relative to an otherwise equivalent LSTM energy-aware policy, while reducing total energy consumption by 24.70% and the number of migrations by 50%. A constant-carbon-intensity counterfactual further shows that temporal carbon-aware scheduling contributes an additional 3.05 percentage points of emissions reduction beyond the reduction attributable to lower energy consumption alone. The migration-energy-weighted carbon intensity is reduced from 269.66 to 174.65 gCO2e/kWh, while average access latency remains essentially unchanged. These results demonstrate that combining workload prediction, migration-cost awareness, and temporally varying carbon signals can improve the environmental efficiency of tiered storage management without compromising system performance.

---