Automatic I/O stream management based on file characteristics

Abstract

Data fragmentation exists widely in SSD NAND flash, which greatly increases garbage collection overhead. Dividing data of different lifetimes into different I/O streams can effectively solve the problem of data fragmentation. Therefore, the management of I/O streams is the key to reduce data fragmentation and thereby improve the lifetime and performance of SSD. In this paper, we propose FileStream, a file-based automatic stream management scheme with machine learning, which can be directly applied to multiple workloads. We find that file characteristics are effective in distinguishing data lifetimes, and thus FileStream divides data into different I/O streams based on the file characteristics. The results of six benchmarks on real SSD show that FileStream reduces the write amplification factor by 34.5% and 21.6% on average, compared to the baseline and two advanced automatic stream management methods, respectively.