SEOUL, November 11 (AJP) - Next-generation flash storage can run up to three times faster with a software fix alone, KAIST researchers have found, after tracing the technology's performance bottleneck to the computer's central processor rather than the storage hardware itself.
KAIST said Wednesday that a team led by Won You-jip, a professor in its School of Electrical Engineering, identified the software chokepoint and developed a new file system technology called GAZE to remove it. A file system is the software that decides where a computer stores files and how it reads and writes them.
The work targets Zoned Namespace, or ZNS, flash storage, a next-generation design that divides storage space into large sections called zones and records data in sequence. The approach lightens the management burden inside the storage device, but it hands the host computer a demanding job in return. To reuse space, the operating system must move the data worth keeping out of a zone and reclaim the emptied area, a routine known as garbage collection.
Existing file systems were built to clean up relatively small areas at a time. On ZNS devices, a single cleanup can involve moving hundreds of megabytes to several gigabytes of data, which stretches cleanup times and delays other data work.
When the team measured the problem, cleaning a one-gigabyte zone took about 1.7 seconds. Most of that burden came not from the storage device but from the central processing unit, which accounted for about 73 percent of the processing load in the read stage and about 84 percent in the write stage as it tracked data status and managed memory. Faster storage hardware alone could not solve the problem.
GAZE reworks the cleanup in three ways. It splits the job of checking, reading and rewriting data into four stages that run at the same time, like a factory production line, instead of one after another. It sets aside dedicated memory for cleanup in advance rather than requesting new memory for every task. And instead of freezing all incoming writes until cleanup finishes, it keeps accepting the write requests that can wait in memory.
The team built GAZE into F2FS, the file system that manages flash storage on the Linux operating system, and tested it under a range of workloads. Compared with the existing ZNS version of F2FS and earlier research prototypes, data throughput rose by as much as three times, and garbage collection time fell to about a third of the baseline.
The technology requires no new dedicated hardware and no change to the format of the data written on the device, which means it can be adopted through software alone. The researchers said its core design principles could also carry over to other log-structured file systems, which record new data in sequential layers.
The gains matter most where storage must feed large volumes of data without pause, including data centers and artificial intelligence inference, the process by which a trained AI model produces answers. However fast a processor runs, its output slows if it sits waiting for data. The study measured file system performance, and whether the technology shortens actual AI response times remains to be verified in deployment.
"For AI to answer quickly, supplying the needed data on time matters as much as computing power," Won said. "We expect this technology to become a foundation for faster AI inference by cutting the time and waiting that storage cleanup consumes."
KAIST postdoctoral researcher Kim Ju-won was the first author of the study, with master's student Yoo Hwan-seok and doctoral student Yoo Seung-won as co-authors and Won as corresponding author. The research was supported by the National Research Foundation of Korea, the Institute of Information and Communications Technology Planning and Evaluation, and SK hynix.
The results were presented on Sept. 30 at ACM SOSP 2026 in Prague, one of the top international conferences in operating systems research.
[Reference Information]
Journal/Source: ACM SOSP 2026
Title: GAZE: Rearchitecting Log-Structured Filesystem for Garbage Collection on Zoned Namespace Flash Storage
Link/DOI: 10.1145/3830418.3843870
Copyright ⓒ Aju Press All rights reserved.