KAIST researchers have developed GAZE, a file-system technology designed to reduce delays when next-generation zoned namespace flash storage reclaims space. The research team, led by Professor Won Youjip of the School of Electrical Engineering, examined why cleaning large storage zones can interrupt applications even when the physical storage device is not the main source of delay.
ZNS flash stores data sequentially in large zones, reducing management work inside the device. However, reclaiming a zone can involve hundreds of megabytes to several gigabytes at once. In an analysis of a 1GB zone, the team found that the process took about 1.7 seconds, with most of the time spent on file-system status checks and memory management by the CPU. CPU work accounted for about 73% of the read-stage burden and 84% of the write-stage burden.
GAZE overlaps data checking, reading, write preparation and writing through pipelining. It also reserves dedicated memory in advance to reduce repeated dynamic allocation, while accepting write requests in memory when immediate storage space is not required. These measures are intended to reduce interruptions during space reclamation rather than eliminate the underlying cleaning process.
Applied to F2FS in Linux 6.1.0, GAZE delivered up to three times the throughput of existing ZNS support methods and earlier techniques. The space-reclamation process took about one-third as long as with conventional F2FS in the evaluation. The findings were presented at ACM SOSP 2026 in Prague on September 30, and the team identified AI inference servers, data centers and high-performance Android devices using F2FS as potential application areas.
