Compression for Tiering: Enabling High-Tier Flash Storage for Hot Data via Cold Data Compression

Abstract

Advances in flash memory technology have increased the number of bits stored per cell, significantly reducing the cost of solid-state drives (SSDs). As a result, SSDs using high-density flash have become common, though they suffer from lower performance and endurance than low-density flash. Modern SSDs exploit the ability of high-density flash to operate in single-level cell (SLC) mode without hardware modification, using part of the flash as an SLC write buffer in a hybrid architecture. However, as space utilization grows, the SLC buffer must shrink to preserve user-visible capacity, increasing migration overhead and degrading performance. Thus, conventional SSDs can use the SLC buffer only sparingly under high-utilization conditions. To overcome this limitation, we propose Compression for Tiering (CFT), a novel Flash Translation Layer (FTL) architecture that selectively compresses highly compressible cold data to reclaim internal space and provisions the reclaimed space as an SLC-mode tier for hot data. Unlike prior schemes that borrow SLC capacity from unused free blocks, CFT provisions the tier from reclaimed internal space, so its size is less tightly coupled to space utilization than in conventional hybrid SSDs. Unlike prior schemes that treat SLC regions merely as transient write buffers, CFT keeps hot data resident in the SLC tier to sustain fast read performance. When workloads provide sufficiently compressible cold data, CFT can retain a larger SLC tier than utilization-only schemes even at high capacity utilization, improving both read and write performance. Evaluations using FIO synthetic workloads and MSR traces show that CFT’s benefit is strongly condition-dependent: the largest improvements—11.3% on synthetic workloads and 10.8% on real traces—arise only under a favorable combination of clear hot/cold separation, high data compressibility, and sufficient idle time, whereas the typical improvement across the tested conditions is considerably smaller, dropping to near zero—or a slight slowdown—when access skew is low or data compress poorly. A sensitivity study over CFT’s threshold parameters and scaled variants of the baseline further indicates that these results are not artifacts of parameter or baseline tuning.