ATIC: a closed-loop, migration-aware tiering controller for cloud and edge NVMe storage

Tiered cloud and edge storage pairs a small, fast tier with a larger capacity tier, but traditional storage replacement policies can allow for the admission of large or sequentially accessed objects whose migration cost is greater than the object’s reuse value. This paper introduces an online admission and eviction controller, ATIC, that combines interpretable benefits per byte of decayed frequency, compactness based on reuse-gap (rather than reuse-size), sequentially, read propensity, object size, device service-time savings, and promotion cost. A quantile gate dynamically tunes the admission selectivity every 750 requests, based on the ratio of promoted bytes to requested bytes. ATIC is benchmarked in a disclosed object-level queueing simulator with five business-intelligence and five artificial-intelligence workload families, 30,000 requests per workload and 20 paired seeds. Relative to GDSF, ATIC saves 41.3%–80.3% of promoted bytes and 2.9%–10.5% of modeled active-I/O energy at 10% capacity for all workloads (one-sided paired Wilcoxon p < 10⁻⁴). With the best latency baseline, ATIC reduces the latency for AI training by 0.5%, and checkpoint traffic by 2.6%, outperforming effectively tied for ETL, and 10.5% and 3.6% slower for locality-dominant BI and mixed analytics, respectively. Ablations reveal that size normalization keeps latency failure from becoming widespread, while the adaptive gate mostly restricts migration. ATIC thus offers a clear migration-energy-latency trade-off, instead of a general latency improvement. Before deploying, hardware and production-trace validation is still required.