A Latency-Aware Garbage Collection Strategy

Abstract

Due to the erase-before-write feature of NAND flash memory, Solid-State Drive (SSD) performs out-of-place update to reduce write latency. However, the invalid pages caused by update request require periodic garbage collection (GC) operations, which interferes with normal user I/O and induces the long-tail latency issue. In addition, high-density flash memory that stores multiple bits in a cell is proposed by flash memory manufacturers to increase the capacity of SSD. Due to the physical structure of flash memory cells, the word-line of high-density flash memory consists of multiple physical pages with different performance and reliability. The chip test results show that the I/O latency of MSB pages is several times than that of LSB pages. The existing GC optimization strategy doesn’t take into account the impact of the target page type in the GC procedure. A basic optimization idea is to migrate valid pages to high-performance LSB pages during the GC procedure to alleviate the long-tail latency problem. However, unrestrained usage of LSB pages can lead to problems such as wasted flash pages. We propose a latency-aware garbage collection strategy for multi-bit NAND flash memory. By monitoring the device queue depth and dynamically adjusting the page type allocated by GC tasks, it can effectively reduce the impact of GC on the overall performance of SSD. The experimental results show that the proposed algorithm reduces the average response time of read requests by 10.9% and the average response time of write requests by 15.8%, respectively. The proposed algorithm can significantly improve the I/O performance of SSD