[论文解读] Nemo: A Low-Write-Amplification Cache for Tiny Objects on Log-Structured Flash Devices

Xufeng Yang, Tingting Tan+8 more|arXiv (Cornell University)|Mar 10, 2026

Advanced Data Storage Technologies被引用 0

一句话总结

Nemo 通过使用小哈希空间和 SG 级分组，将集合关联的闪存缓存重新设计为处理微小对象，以最小化应用层写放大，同时采用近似索引和选择性元数据下移来维持内存效率。

ABSTRACT

Modern storage systems predominantly use flash-based SSDs as a cache layer due to their favorable performance and cost efficiency. However, in tiny-object workloads, existing flash cache designs still suffer from high write amplification. Even when deploying advanced log-structured flash devices (e.g., Zoned Namespace SSDs and Flexible Data Placement SSDs) with low device-level write amplification, application-level write amplification still dominates. This work proposes Nemo, which enhances set-associative cache design by increasing hash collision probability to improve set fill rate, thereby reducing application-level write amplification. To satisfy caching requirements, including high memory efficiency and low miss ratio, we introduce a bloom filter-based indexing mechanism that significantly reduces memory overhead, and adopt a hybrid hotness tracking to achieve low miss ratio without losing memory efficiency. Experimental results show that Nemo simultaneously achieves three key objectives for flash cache: low write amplification, high memory efficiency, and low miss ratio.

研究动机与目标

• Motivate and quantify the high application-level write amplification (ALWA) in tiny-object flash caches and identify root causes.
• Propose Nemo as a cache architecture that achieves near-ideal WA while preserving memory efficiency and caching performance.
• Design Nemo components: small hash-space set-group (SG), SG-level flushing/eviction, PBFG-based approximate indexing, and selective in-memory/offloaded metadata.
• Demonstrate that Nemo reduces ALWA and DLWA while maintaining acceptable memory overhead and performance.

提出的方法

• Analyze the sources of write amplification in hierarchical flash caches and model L2SWA (log-to-set write amplification) components.
• Introduce Nemo’s architecture with a small in-memory hash space organized as Set-Groups (SGs) and SG-level batched writes to on-flash SG pools.
• Employ Parallel Bloom Filter Group (PBFG) for approximate object indexing to reduce memory footprint and facilitate parallel lookups.
• Use selective metadata offloading (keeping hot metadata in memory, offloading stable metadata to flash) and a hybrid 1-bit hotness tracker to minimize memory overhead.
• Define write amplification of Nemo as WA(Nemo) = 1 / E[FR_SG] and provide rationale that SG-level batching yields high fill rates.
• Outline three design challenges (C1–C3) and corresponding solutions to ensure well-filled SGs, efficient offloading, and effective eviction.

Figure 1 . Application-level write amplification comparison. “S” denotes Set, “SG” denotes Set-Group.

实验结果

研究问题

• RQ1 What are the major sources of write amplification in existing tiny-object flash caches (e.g., FairyWREN, Kangaroo) and how can they be mitigated?
• RQ2 Can a set-group based, small-hash-space design with SG-level batching reduce ALWA to near-ideal levels without sacrificing cache performance or memory overhead?
• RQ3 How can approximate indexing (PBFG) and selective metadata offloading sustain memory efficiency while supporting fast lookups for tiny objects?
• RQ4 What is the impact of hotness tracking strategy on memory footprint and cache effectiveness in Nemo?

主要发现

• Nemo can achieve ALWA close to the theoretical optimum (1.56) with DLWA as low as 1 on existing log-structured SSDs.
• Nemo reduces flash writes by up to 90% while maintaining low memory overhead (about 8.3 bits per object for metadata) compared to prior designs.
• PBFG indexing enables in-memory indexing with around 9.6 bits per object at 1% false positive rate, significantly reducing memory cost.
• SG-level batching and a FIFO on-flash SG pool align flush/eviction with SSD log-structured behavior to minimize write amplification.
• Approximate indexing and hybrid hotness tracking preserve caching performance and memory efficiency, enabling practical deployment within CacheLib.
• Compared to log-structured front tiers or traditional set-associative designs, Nemo eliminates log-to-set migration overhead and mitigates RMW-related amplification.

Figure 2 . Hierarchical cache.