KAIST's GAZE File System Lifts Flash Storage Throughput Up to Threefold

By Amara Osei

A file system developed at the Korea Advanced Institute of Science and Technology (KAIST) raises data processing performance on next-generation ZNS flash storage by up to three times, the university announced on the 11th. The system, called GAZE, requires no new hardware and no change to on-device data formats.

Professor Youjip Won's team at the School of Electrical Engineering identified a software bottleneck in ZNS (Zoned Namespace) flash storage and rebuilt the file system's cleanup logic around it.

Why does cleanup slow flash storage down?

A processor can only work as fast as the storage device feeds it data. However fast the chip, processing stalls when it waits on the digital warehouse below it.

Flash storage also has to tidy itself while in use. The device moves data that must be kept to new locations and reclaims the emptied space — a process called garbage collection, comparable to shifting stock around a warehouse to free room for new goods.

ZNS technology divides storage into large sections called zones and writes data sequentially. Flash storage is semiconductor-based and retains data without power; managing it zone by zone reduces the device's complex internal operations. The trade-off sits in cleanup: existing file systems were designed to reclaim relatively small areas, but a single ZNS zone can hold several gigabytes.

The team's measurements quantify the gap. Under the existing approach, reclaiming a 1-gigabyte zone took about 1.7 seconds.

What does GAZE change?

Won's team rebuilt the cleanup method along three lines:

• It split the process of checking, reading and writing data into four stages that run simultaneously.
• It secured the memory needed for cleanup in advance.
• It kept accepting data that could first be held in memory during cleanup, replacing a scheme that made every storage request wait until cleanup finished.

The researchers implemented GAZE on F2FS, the file system Linux uses to manage files on flash storage, and benchmarked it under various workloads. Against stock F2FS with ZNS, and against earlier technologies ZNS+ and P-GC, the volume of data processed in a given time rose by up to three times. Garbage-collection time fell to roughly one-third of the existing F2FS figure.

The gain comes without dedicated hardware and without altering the format of data written to the device — the feature the team singles out as the technology's distinguishing property.

Where does it matter — and what is unproven?

The target application is AI inference in data centers. "Improvements like these can help use storage devices more efficiently in AI inference that handles large volumes of data and in data centers," the team said, "because reducing the time and processing burden consumed by data cleanup leaves room to supply the data needed for AI computation more smoothly."

The team also flagged the limit: the actual effect on shortening AI inference response times remains a task requiring future verification.

Won framed the problem in supply terms: "For AI to answer quickly, it is important not only to have computing power but also to supply the necessary data in a timely manner. We expect this technology to become a foundation for speeding up AI inference by reducing the time and waiting involved in organizing data in storage devices."

The research was presented on Sept. 30 at ACM SOSP 2026, a leading international conference on operating systems, held in Prague, Czech Republic.

A 1.7-second cleanup per gigabyte, cut to a third, is the number data-center operators will test first.