Researchers at the Korea Advanced Institute of Science and Technology have developed GAZE, a filesystem technology designed to eliminate software bottlenecks in next-generation flash storage devices and significantly accelerate data processing. Led by Professor Youjip Won from the KAIST School of Electrical Engineering, the engineering team identified that inefficient space management software frequently prevents high-performance storage media from operating at full potential. By restructuring the internal software procedures that relocate and reorganize stored information during space reclamation, the newly developed system achieves up to three times higher data-processing performance compared to existing zoned storage management systems.
Flash storage operates as semiconductor-based nonvolatile memory that retains stored records even when electrical power is disconnected, serving as a core foundation for modern solid-state drives. Storage devices play a crucial role during artificial intelligence inference, which is the operational process where an AI model generates responses or outputs based on its previously learned parameters. Because an artificial intelligence workflow requires timely access to model weights and input information, system efficiency can degrade substantially if high-speed computational processing units are forced to idle while waiting for data retrieval from underlying storage arrays.
The research investigation centered specifically on the space cleanup mechanism operating within flash storage hardware, commonly designated as garbage collection. To reuse physical blocks that have previously been written, a flash storage device must copy active information that must be retained into an alternative physical location before erasing and reclaiming the vacated region. This operational cycle functions analogously to shifting essential warehouse inventory into consolidated storage bays so that empty floor sections can be repurposed for incoming deliveries.
The experimental evaluation focused on Zoned Namespace flash storage, an architecture that organizes physical memory into expansive partitions called zones and enforces sequential data writes within each individual zone. While this structural partition reduces the complexity of internal drive controllers by delegating storage layout decisions upward, it places a heavy administrative burden onto the host operating system and filesystem. Operating systems must assume direct responsibility for managing and orchestrating cleanup across these substantial storage zones.
Conventional filesystems were originally architected to clean comparatively compact storage sectors. In contrast, Zoned Namespace storage devices routinely require the migration of several gigabytes of accumulated data during the evacuation of a single zone. Although the physical volume of data requiring migration grew dramatically under zoned architectures, the underlying operational procedures remained largely unaltered, introducing severe latency overheads into system input-output pipelines.
Quantitative analysis conducted by the research group revealed that cleaning a single one-gigabyte zone using conventional filesystem methods required approximately 1.7 seconds to complete. The investigators demonstrated that a substantial fraction of the observed processing delay during data reading and rewriting stemmed from host central processing unit activity rather than hardware limits of the storage media itself. Specifically, operating system state tracking and repeated system memory allocation during cleanup operations generated heavy central processing unit overhead, demonstrating that faster storage hardware alone could not overcome the bottleneck.
To eliminate these processing delays, the KAIST team re-engineered the garbage collection workflow into three complementary software architectural enhancements. The first modification decomposed the space reclamation sequence into four distinct operational phases: checking data validity, reading existing records, processing memory structures, and rewriting valid records back to storage. Rather than executing these stages sequentially and waiting for each step to finish, GAZE executes the phases concurrently in an overlapped pipeline, reading upcoming data segments while concurrently processing and writing previously retrieved data, mirroring an industrial factory assembly line.
The second structural intervention pre-allocates and reserves dedicated host memory buffers specifically designated for ongoing cleanup operations. This architectural choice eliminates the recurring runtime overhead of repeatedly allocating and freeing temporary operating system memory buffers whenever a zone cleanup triggers. Third, the GAZE filesystem continues accepting new write requests that can be temporarily buffered in memory without immediately requiring physical disk allocation while active zone cleanup is executing, overcoming legacy limitations where write tasks were forced into idle queues until cleanup finished.
The research team implemented the GAZE architecture directly within the Flash-Friendly File System on the Linux 6.1.0 kernel environment and evaluated its behavior across a spectrum of workloads. When benchmarked against the standard Zoned Namespace implementation in baseline F2FS as well as earlier specialized technologies including ZNS+ and P-GC, GAZE delivered up to a three-fold improvement in overall data-processing throughput. Furthermore, the operational time required to complete space cleanup was reduced to approximately one-third of the duration recorded by standard F2FS.
A notable characteristic of the GAZE architecture is that it delivers these measured speed improvements entirely through operating system software optimization, requiring no proprietary semiconductor modifications or revisions to physical storage recording formats. The researchers demonstrated that the core algorithmic principles governing GAZE can be generalized across other log-structured filesystems, which uniformly record incoming write streams sequentially to optimize storage endurance and efficiency.
By streamlining the software data path and eliminating idle states, the filesystem provides a foundation to support data-intensive environments including artificial intelligence inference pipelines, enterprise cloud data centers, and high-performance computing clusters. First author Dr. Juwon Kim of the KAIST Institute for Information Technology Convergence conducted the study alongside co-authors Seungwon Yoo and Hwanseok Yoo, with Professor Youjip Won serving as corresponding author. The peer-reviewed research was presented on September 30 at the ACM Symposium on Operating Systems Principles in Prague, Czech Republic, supported by the National Research Foundation of Korea, the Institute of Information & Communications Technology Planning & Evaluation under South Korea’s Ministry of Science and ICT, and SK hynix.
Subject of Research: Filesystem architecture for zoned namespace flash storage garbage collection
Article Title: KAIST removes “speed bottleneck” in flash storage… expected to accelerate AI inference
Article References: KAIST removes “speed bottleneck” in flash storage… expected to accelerate AI inference. Original publication (Bibliographic details unavailable.)
Image Credits: AI Generated
DOI: Not provided
Keywords: GAZE, flash storage, filesystem architecture, garbage collection, zoned namespace storage, ZNS, F2FS, Linux 6.1.0, SSDs, AI inference, KAIST, Youjip Won, log-structured filesystem, data centers, ACM SOSP 2026, CPU memory management
News Source: Denise Maddox. (November 11, 2026). KAIST Filesystem GAZE Accelerates Flash Storage by Overhauling Garbage Collection Software. Scienmag.
