11-19 更新 实时
首页/ 算力/ 正文

研究者消除闪存存储速度瓶颈以加速 AI 推理

算力 来源:TechXplore 2026-11-10 23:00 0 阅读 机器翻译
研究者消除闪存存储速度瓶颈以加速 AI 推理

由 Lisa Lock 编辑,Andrew Zinin 审阅

本文已按照 Science X 的编辑流程与政策进行审阅。编辑在确保内容可信度的同时,重点标注了以下属性:

添加到首选来源

在垃圾回收期间处理写请求:传统方法(左)与 GAZE(右)。图片来源:韩国科学技术院(KAIST)

即便使用高性能存储设备,低效的数据管理也会使其无法发挥出全部速度。KAIST 的研究人员开发出一项技术,改进了数据的迁移与重组过程以复用存储空间,从而提升了下一代闪存存储设备的数据处理性能。

由电气工程学院教授 Youjip Won 领导的团队找出了拖慢下一代闪存存储设备数据处理的软件瓶颈,并开发出文件系统技术「GAZE」,改进了存储空间的管理方式。文件系统是一种软件,负责管理计算机将文件存放在何处,以及如何读写这些文件。该论文发表于 Proceedings of the ACM SIGOPS 32nd Symposium on Operating Systems Principles。

闪存是基于半导体的存储,即使断电也能保留数据,被用于计算机 SSD 等设备。存储设备在「推理」过程中同样扮演着重要角色——推理即人工智能基于所学内容生成答案的过程——因为所需的数据必须在恰当的时间送达。即使计算单元速度很快,如果等待数据的时间过长,整个系统的效率也会下降。

研究团队聚焦于存储设备内部的“空间清理”过程。为了重新利用此前被占用的空间,闪存会把必须保留的数据转移到另一位置,并回收腾出的区域。这一过程被称为“垃圾回收”。可以把它比作搬运必要物品并重新整理仓库的各个区域,从而让空出的空间能够再次被使用。

传统的垃圾回收过程。图片来源:韩国科学技术院(KAIST)

GAZE 的流水线式垃圾回收。图片来源:韩国科学技术院(KAIST)

本研究考察的 ZNS 闪存存储将存储空间划分为称为“zone”的大区域,并在每个 zone 内按顺序写入数据。这可以减少存储设备内部复杂的管理操作,但要求操作系统和文件系统能够高效地处理这些大 zone 的清理。

现有的文件系统主要针对清理相对较小的区域而设计。然而在 ZNS 存储中,清理单个 zone 时可能需要移动数 GB 的数据。清理任务的规模变大了,而执行清理的方法却基本未变,结果导致延迟变长。

研究团队的分析显示,使用传统方法清理一个 1 GB 的 zone 大约需要 1.7 秒。读取和重写数据所产生的处理开销中有相当一部分并非来自存储设备本身,而是来自中央处理器(CPU)执行的状态管理和内存管理操作。这意味着,仅仅让存储设备变得更快无法解决这一问题。

为解决这一问题,研究人员从三个方面重新设计了清理流程。首先,他们把数据的检查、读取、处理和重写工作划分为四个可以相互重叠的阶段。系统不再等待一项任务完全结束,而是在处理并写入此前读取的数据的同时,读取下一批数据。其原理类似于工厂中多道工序同时运转的生产线。

其次,系统会预先为清理操作保留专用内存。这减少了每次执行该操作时反复分配和管理临时内存所带来的开销。第三,在清理进行期间,系统继续接受那些并不立即需要额外存储空间的写入请求。这改进了此前的做法——在那种方式下,即便是可以先存入内存的数据,也必须等到清理过程完成。

研究团队在 Linux 6.1.0 的 F2FS 文件系统中实现了该技术,并比较了其在各种工作负载下的性能。与 F2FS 常规的 ZNS 支持以及 ZNS+、P-GC 等此前技术相比,GAZE 的数据处理性能最高提升至三倍。空间清理所需的时间也降至常规 F2FS 的大约三分之一。

该技术的另一个关键特点是,它无需额外的专用硬件,也无需改变数据写入存储设备的格式,就能提升性能。相反,研究人员改进了现有文件系统软件执行操作的方式,并展示了将核心设计原则应用于其他日志结构文件系统的潜力。日志结构文件系统会按数据到达的顺序依次记录新数据。

该研究表明,要充分利用存储设备的性能,不仅需要改进半导体硬件,还需要改进管理硬件的软件。通过减少数据路径上不必要的等待和管理操作,研究人员得以将高性能存储设备的速度转化为实际的数据处理性能。

在 AI 推理、数据中心以及处理海量数据的高性能计算环境中,存储设备能否顺畅、高效地提供所需数据至关重要。这项新技术有望成为提升此类环境中数据处理效率的基础。

在本研究所展示的文件系统性能改进的基础上,将该技术扩展到更广泛的存储设备和工作负载,并验证其在 AI 推理环境中的数据供给效率,或可进一步明确其在真实服务中的实用价值。

Won 表示:“要让韩国的先进存储技术充分发挥性能,消除相关的软件瓶颈至关重要。这项研究意义重大,因为我们彻底消除了阻碍快速闪存存储设备被充分利用的操作系统瓶颈。”

“通过将其应用扩展至 AI 推理以及处理大规模数据的数据中心,我们期望能为提升高性能存储的利用效率以及工业数据处理的竞争力做出贡献。”

Juwon Kim 等,GAZE: Rearchitecting Log-Structured Filesystem for Garbage Collection on Zoned Namespace Flash Storage,ACM SIGOPS 第 32 届操作系统原理研讨会论文集(2026)。DOI: 10.1145/3830418.3843870

这篇报道的作者是谁?

艺术史学士,物质文化硕士。曾任博物馆编辑、急救员和器官移植协调员。自 2021 年起为 Science X 担任编辑。完整简介 →

物理学硕士,具备科研经历。长期热衷科学新闻。在 Science X 的编辑工作中发挥关键作用。完整简介 →

引用:研究人员消除闪存存储中的“速度瓶颈”以助力加速 AI 推理(2026 年 11 月 10 日),2026 年 11 月 10 日检索自 https://techxplore.com/news/2026-11-bottleneck-storage-ai-inference.html

本文受版权保护。除为私人学习或研究目的进行的合理使用外,未经书面许可不得以任何形式复制本文任何部分。内容仅供参考。

获取即时摘要文本(Gist)

GAZE 利用流水线处理、预分配内存和选择性写入处理,加速 ZNS 闪存存储中的垃圾回收。它在 Linux F2FS 中实现,吞吐量最高提升至三倍,并将清理时间降至传统 F2FS 的大约三分之一,且无需额外硬件或磁盘格式更改。

本摘要由 LLM 自动生成。完整免责声明

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Researchers remove 'speed bottleneck' in flash storage to help accelerate AI inference)

edited by Lisa Lock, reviewed by Andrew Zinin

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:

Add to Preferred Sources

KAIST Removes "Speed Bottleneck" in Flash Storage… Expected to Accelerate AI Inference Handling write requests during garbage collection: The conventional method (left) and GAZE (right). Credit: The Korea Advanced Institute of Science and Technology (KAIST)

Even with high-performance storage devices, inefficient data management can prevent them from reaching their full speed. KAIST researchers have developed a technology that improves the process of moving and reorganizing data to reuse storage space, increasing the data-processing performance of next-generation flash storage devices.

The team led by professor Youjip Won of the School of Electrical Engineering identified the software bottleneck that slows data processing in next-generation flash storage devices and developed "GAZE," a filesystem technology that improves the way storage space is managed. A filesystem is software that manages where a computer stores files and how it reads and writes them. The paper is published in the Proceedings of the ACM SIGOPS 32nd Symposium on Operating Systems Principles.

Flash storage is semiconductor-based storage that retains data even when power is turned off and is used in devices such as computer SSDs. Storage devices also play an important role in "inference," the process in which artificial intelligence generates answers based on what it has learned, because the required data must be supplied at the right time. Even if the computing unit is fast, the efficiency of the entire system can decline if too much time is spent waiting for data.

The research team focused on the "space cleanup" process inside storage devices. To reuse previously occupied space, flash storage moves data that must be retained to another location and reclaims the vacated area. This process is known as "garbage collection." It can be compared to moving necessary items and reorganizing sections of a warehouse so that empty space can be used again.

KAIST Removes "Speed Bottleneck" in Flash Storage… Expected to Accelerate AI Inference The conventional garbage collection process. Credit: The Korea Advanced Institute of Science and Technology (KAIST)

KAIST Removes "Speed Bottleneck" in Flash Storage… Expected to Accelerate AI Inference GAZE’s pipelined garbage collection. Credit: The Korea Advanced Institute of Science and Technology (KAIST)

The ZNS flash storage examined in this study divides storage space into large areas called "zones" and writes data sequentially within each zone. This can reduce complex management operations inside the storage device, but it requires the operating system and filesystem to efficiently handle cleanup of these large zones.

Existing filesystems were designed primarily to clean relatively small areas. In ZNS storage, however, several gigabytes (GB) of data may need to be moved when a single zone is cleaned. The scale of the cleanup task increased while the method of performing it remained largely unchanged, resulting in longer delays.

The research team's analysis showed that cleaning a 1 GB zone using the conventional method took approximately 1.7 seconds. A substantial portion of the processing overhead generated while reading and rewriting data came not from the storage device itself but from state management and memory management operations performed by the central processing unit (CPU). This meant that simply making the storage device faster could not solve the problem.

To address this issue, the researchers redesigned the cleanup process in three ways. First, they divided the work of checking, reading, processing and rewriting data into four stages that could overlap with one another. Instead of waiting for one task to finish completely, the system processes and writes previously read data while simultaneously reading the next set of data. The principle is similar to a production line in which multiple stages of a factory operate at the same time.

Second, the system reserves dedicated memory for cleanup operations in advance. This reduces the overhead of repeatedly allocating and managing temporary memory each time the operation is performed. Third, while cleanup is underway, the system continues accepting write requests that do not immediately require additional storage space. This improves on the previous approach, in which even data that could first be stored in memory had to wait until the cleanup process was completed.

The research team implemented the technology in the F2FS filesystem on Linux 6.1.0 and compared its performance under various workloads. Compared with the conventional ZNS support of F2FS and previous technologies such as ZNS+ and P-GC, GAZE achieved up to three times higher data-processing performance. The time required for space cleanup was also reduced to approximately one-third of that of conventional F2FS.

Another key feature of the technology is that it does not require additional dedicated hardware or changes to the format in which data are written to the storage device in order to improve performance. Instead, the researchers improved the way existing filesystem software performs its operations and demonstrated the potential to apply the core design principles to other log-structured filesystems. A log-structured filesystem records new data sequentially as it arrives.

The study demonstrates that fully utilizing the performance of storage devices requires improvements not only in semiconductor hardware but also in the software that manages it. By reducing unnecessary waiting and management operations along the data path, the researchers were able to translate the speed of high-performance storage devices into actual data-processing performance.

In AI inference, data centers and high-performance computing environments that handle large volumes of data, it is important for storage devices to supply required data smoothly and efficiently. The new technology is expected to serve as a foundation for improving data-processing efficiency in such environments.

Building on the filesystem performance improvements demonstrated in this study, expanding the technology to a wider range of storage devices and workloads and verifying data-supply efficiency in AI inference environments could further clarify its practical value for real-world services.

Won said, "For Korea's advanced memory technologies to deliver their full performance, it is essential to eliminate the associated software bottlenecks. This study is significant because we completely removed the operating-system bottleneck that prevented fast flash storage devices from being fully utilized.

"By expanding its application to AI inference and data centers that handle large-scale data, we expect to contribute to improving both the utilization efficiency of high-performance storage and the competitiveness of industrial data processing."

Juwon Kim et al., GAZE: Rearchitecting Log-Structured Filesystem for Garbage Collection on Zoned Namespace Flash Storage, Proceedings of the ACM SIGOPS 32nd Symposium on Operating Systems Principles (2026). DOI: 10.1145/3830418.3843870

Who's behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →

Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →

Citation: Researchers remove 'speed bottleneck' in flash storage to help accelerate AI inference (2026, November 10) retrieved 10 November 2026 from https://techxplore.com/news/2026-11-bottleneck-storage-ai-inference.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.

Get Instant Summarized Text (Gist)

GAZE accelerates garbage collection in ZNS flash storage using pipelined processing, preallocated memory and selective write handling. Implemented in Linux F2FS, it achieved up to threefold higher throughput and reduced cleanup time to about one-third of conventional F2FS, without additional hardware or on-disk format changes.

This summary was automatically generated using LLM. Full disclaimer

This site uses cookies to assist with navigation, analyse your use of our services, collect data for ads personalisation and provide content from third parties. By using our site, you acknowledge that you have read and understand our Privacy Policy and Terms of Use.

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 TechXplore(Researchers remove 'speed bottleneck' in flash storage to help accelerate AI inference),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗