Make Lemonade out of Lemons: A Learning-based ZNS SSDs with Built-in Compression

Zoned Namespace (ZNS) SSDs offer a promising approach to simplifying flash management and reducing storage costs. Enabling built-in compression in ZNS SSDs can further enhance storage efficiency and improve performance without increasing the host’s burden. However, compression introduces variable-size data, disrupting the alignment between logical and physical page addresses. This misalignment, together with ZNS SSDs’ small internal DRAM resources, introduces three fundamental challenges: (1) efficiently locating the physical address of compressed data at low memory usage, (2) deterministically addressing multiple compressed logical pages packed within a physical page without maintaining large per-logical-page metadata, and (3) avoiding severe performance degradation caused by serialized reads when compressed data spans multiple physical pages without incurring large memory overhead. To address these challenges, we propose Lemonade , a ZNS SSD design with built-in compression that enables efficient compact storage while maintaining low memory overhead. The design is based on three key insights driven by the unique characteristics of ZNS SSDs. First, we observe that the sequential write and append-only constraints of ZNS naturally yield an approximately linear LPN-PPN mapping after compression. Lemonade exploits this by employing lightweight learned linear models, enabling physical address lookup with low memory cost (8 bytes per model). Second, in ZNS SSDs, compressed logical pages packed within a physical page have consecutive logical page numbers, and each expands to a fixed-size logical page after decompression. Lemonade exploits this to eliminate heavy metadata for intra-page offsets. Instead, it determines the location of compressed data by small metadata and simple arithmetic. Third, Lemonade mitigates the read latency penalty caused by accessing cross-page data through a scheduling strategy that overlaps the corresponding read requests. This strategy does not introduce additional memory overhead, as it exploits the ZNS SSD’s physical layout where physical pages containing cross-page data are typically placed on different channels and dies. We implement Lemonade on the FEMU ZNS SSD emulator and evaluate it with real-world datasets and storage workloads. Experimental results demonstrate that Lemonade achieves both compact storage and high performance while maintaining low DRAM usage. Lemonade delivers up to 51.1% higher write throughput and 22.2% higher read throughput than the state-of-the-art design. Meanwhile, it significantly reduces the storage footprint and requires only 8 bytes of DRAM per zone for indexing in the best case.