A lightweight fine-grained scheme for distinguishing the hotness of warm data to reduce segment cleaning overhead
Lihua Yang, Yang Xiao, Zhipeng Tan, Fang Wang, Weizhao Lin, Wei Zhang, Jiaxin Li, Kai Lü · Journal of Parallel and Distributed Computing · 2025
With the widespread adoption of flash memory, the Flash Friendly File System (F2FS) designed to flash memory characteristics has become widely-used in large data centers. However, F2FS encounters from significant cleaning overheads due to its logging scheme writes. We observe that warm data in F2FS account for a substantial proportion, at least 80%. Nevertheless, the mixed storage of warm data with varying hotness exacerbates segment cleaning challenges. To address this issue, we propose a scheme called M2H, which involves a fine-grained management of warm data hotness identified by the K-means clustering algorithm. M2H determines hotness by considering factors such as file block update distance, most recently used distance, and workload characteristics. M2H facilitates M ulti-log delayed writing and M odified segment cleaning based on H otness. To reduce costs associated with distinguishing data hotness at the file block level, we employ Mini Batch K-means, which is referred to as HMBK. Moreover, for servers equipped with GPUs, the clustering process can be offloaded to the GPU, known as HGPU. We conduct a comprehensive comparison of traditional F2FS, M2H, HMBK, and HGPU on a real platform. Results show that compared to traditional F2FS, HGPU reduces the number of segment cleanings by 54.41% to 97.93%.