InDeF: An Advanced Defragmenter Supporting Migration Offloading on ZNS SSD
Wenjie Qi, Zhipeng Tan, Jicheng Shao, Lihua Yang, Yang Xiao
Abstract
The Zoned Namespace (ZNS) emerges as a new storage interface with advantages such as low garbage collection overhead and low over-provisioning cost. However, it is vulnerable to fragmentation due to its out-of-place updates and the multi-threaded writing behaviors of applications. Fragmentation splits I/O requests and causes resource conflicts within the device, degrading I/O performance. Most defragmentation tools need to read the entire contents of the file from the SSD to the host memory and then rewrite the data to contiguous space in the SSD. The host-level migration operations prolong the elapsed time of defragmentation and excessive write operations for data migration even reduce the device lifetime.Our experiments discover that defragmenting data with a low degree of fragmentation or cold data provides little performance gain. With the observation, we propose a new defragmentation tool called InDeF to reduce the defragmentation overhead. InDeF combines the degree of logical and physical fragmentation and access hotness to filter out the fragments that have little impact on I/O performance for reducing the write traffic of the SSD. To take advantage of the internal flash chip parallelism, InDeF offloads the data migration on the SSD, decreasing the elapsed time of defragmentation. Evaluation results show that InDeF decreases the elapsed time of defragmentation by 91.6%∼94.2% and the amount of data migration by 38.1%∼66.7% compared with conventional defragmentation tools.