Unlocking ZNSSD Parallelism: Reducing Read Tail Latency with Large Zones

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98415

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98415

Title:	Unlocking ZNSSD Parallelism: Reducing Read Tail Latency with Large Zones
Authors:	楊雨潔;Yang, Yu-Chieh
Contributors:	資訊工程學系
Keywords:	ZNS SSD;鍵值儲存;平行度;長尾延遲;Rocksdb;ZenFS;ZNS SSD;Key-Value store;Zoned Namespace SSD;Rocksdb;ZenFS;parallelism;tail latency
Date:	2025-07-29
Issue Date:	2025-10-17 12:45:34 (UTC+8)
Publisher:	國立中央大學
Abstract:	Zone Namespace SSD 透過順序寫入的限制，為現代儲存系統在效能與壽命優化上帶來新契機。然而，現有整合 ZNS 的資料庫（如 RocksDB）在實作上常面臨跨 dies 間讀取量不均的問題，導致系統效能下降與讀取延遲尾端表現惡化。儘管先前研究多聚焦於 compaction 機制與寫入平行化的探討，對於 die 讀取平行性卻著墨甚少。本研究提出一套 compaction 策略及 zone 管理架構，旨在提升讀取平行性並緩解 dies 間的讀取集中問題。系統根據各 die 的即時 I/O 佇列長度，將 zone 分為高、中、低與未定義四類。為因應動態工作負載變化，我們設計了「未定義佇列機制」，將熱度變化不明的 zone 暫時隔離，以避免過早重複使用，並提高 zone 分類的準確性。此外，我們提出主動式與被動式 compaction 策略，能夠針對高負載 dies 的熱門資料進行選擇性搬移，將其移至負載較低的 dies，同時嚴格避免跨佇列寫入，以維持負載隔離效果。為支援該策略，系統也整合了一套 zone 分配邏輯，能在必要時主動將 zone 從過載 dies 移出。最後，我們實作了 zone 管理機制，優先關閉高負載佇列中使用率較高的 zone，以達成較高的區域利用率。實驗結果顯示，無論在 db_bench 或實際 IO trace（SNIA、Alibaba）中，所提方法皆能顯著改善 die 的讀取集中情形，達成降低平均延遲及緩解長尾延遲的目的。 ;Zone Namespace Solid-State Drives (ZNS SSDs) offer new opportunities for performance and longevity optimization in modern storage systems by enforcing sequential write constraints within zones. However, existing ZNS-integrated key value stores such as RocksDB with ZenFS often suffer from imbalanced read traffic across dies, leading to degraded performance and significant tail latency. While previous studies have explored compaction and write parallelism, little attention has been paid to die-level read parallelism—a crucial factor affecting latency-sensitive workloads. In this work, we propose a queue-based compaction and zone management framework designed to enhance read parallelism and mitigate load concentration across dies. Our system classifies zones into four queues—high, medium, low, and undefined—based on the real-time I/O queue lengths of their associated dies. To adapt to dynamic workload patterns, we introduce an undefined queue mechanism that isolates zones with uncertain post-compaction heat levels, preventing premature reuse and promoting accurate reclassification. Furthermore, we design proactive and reactive compaction strategies that selectively relocate hot data from congested zones to underutilized ones, while strictly avoiding cross-queue writes to preserve load isolation. Our zone allocation logic further supports this by actively offloading zones from overburdened dies when necessary. Lastly, we implement an active zone management mechanism that prioritizes closing highly utilized zones in high-load queues, maximizing write efficiency while retaining valuable read locality. Experimental results on both synthetic (db_bench) and real world traces (SNIA, Alibaba) demonstrate that our approach significantly improves die-level read distribution, reduces average latency, and effectively mitigates long-tail behavior compared to the baseline ZenFS configuration.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	11	View/Open

社群 sharing

Loading...