Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a technology that can increase data processing performance of next-generation flash storage devices by up to threefold. It is expected to improve the efficiency of AI inference and data processing at data centres by reducing the time needed to clean up storage space.

KAIST said on Tuesday that a team led by Professor Yoo-jip Won (원유집) of the School of Electrical Engineering developed file system technology called GAZE that improves software bottlenecks in next-generation flash storage devices.

A file system is software that manages where a computer stores files and how it reads and writes them. Even if the storage device itself is fast, overall performance can decline if the file system operates inefficiently.

The team focused on next-generation zoned namespace (ZNS) flash storage devices that divide storage space into large sections called zones and record data sequentially.

Flash storage devices perform garbage collection to reuse used space by moving needed data elsewhere and securing free space. Existing file systems are designed to clean relatively small areas, creating a problem in that it takes a long time to handle large sections that run to several gigabytes.

The team’s analysis showed the existing method took about 1.7 seconds to clean a 1 GB section. CPU work to check data processing status and manage memory needed for the task emerged as a key bottleneck.

The team improved the cleanup method in three ways. It first split data checking and reading, preparation for recording, and writing into four stages and ran them in parallel. The approach reduces waiting time by overlapping multiple tasks.

It also secured in advance the memory needed for cleanup to reduce CPU burden from repeated memory allocation. During cleanup, it continued to accept data that could be stored in memory first, minimising unnecessary pauses in work.

The team verified performance by applying GAZE to the F2FS file system on the Linux operating system. Comparing it with the existing F2FS ZNS support method and earlier technologies ZNS+ and P-GC, it found data processing performance improved by up to threefold. The time required for storage cleanup was also shortened to about one-third of that of the existing F2FS.

The technology improves performance through software enhancements alone, without adding dedicated hardware or changing the storage device’s data format. The team expects the technology can be used for AI inference and data centres that handle large-scale data, as well as high-performance computing.

But it said the research confirmed file system performance improvements, and that additional verification is needed on any effect in improving response times for actual AI inference.

Won said, "For AI to answer quickly, it is important not only to have computing power but also to supply the needed data in time." He added, "We expect this technology will become a foundation for speeding up AI inference by reducing the time and waiting involved in storage-device data cleanup."

The research was conducted with support from the National Research Foundation of Korea, the Institute of Information & Communications Technology Planning & Evaluation (IITP) under the Ministry of Science and ICT, and SK Hynix, among others. Joo-won Kim (김주원), a Ph.D. at the KAIST Information & Electronics Research Institute, participated as first author, with Hwan-seok Yoo (유환석), a master’s student, and Seung-won Yoo (유승원), a Ph.D. student, as co-authors. The results were presented on Sept. 30 at the international operating systems conference ACM SOSP 2026, held in Prague, Czech Republic.