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.
KAIST (President Choongsik Bae) announced on November 11 that a research 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.
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 (AI) 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 it spends too much time 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.
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.
Professor Youjip 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,” adding, “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.”
Dr. Juwon Kim of KAIST participated in the study as the first author, while master’s student Hwanseok Yoo and Ph.D. student Seung Won Yoo participated as co-authors. Professor Youjip Won of the KAIST School of Electrical Engineering participated as the corresponding author. The research results were presented on September 30 at ACM SOSP 2026, a major international conference in the field of operating systems held in Prague, Czech Republic.
※ Paper title: GAZE: Rearchitecting Log-Structured Filesystem for Garbage Collection on Zoned Namespace Flash Storage / DOI: 10.1145/3830418.3843870
※ Key author information: First author (Dr. Juwon Kim, KAIST Institute for Information Technology Convergence), co-authors (Seungwon Yoo, Ph.D. student at KAIST; Hwanseok Yoo, master’s student at KAIST), corresponding author (Youjip Won)
This research was supported by the National Research Foundation of Korea project “Research on Order-Preserving Distributed Storage Systems,” the Institute of Information & Communications Technology Planning & Evaluation (IITP) project “Development of Energy-Aware Operating System Technology Based on Disaggregated Architecture,” funded by the Ministry of Science and ICT, and SK hynix, among others.
GAZE: Rearchitecting Log-Structured Filesystem for Garbage Collection on Zoned Namespace Flash Storage
30-Sep-2026