A closed-form expression for write amplification in NAND Flash

Abstract

The log-structured filesystems typically used in current solid-state drive's (SSD) exhibit write amplification, whereby multiple NAND writes are required for each host write. Write amplification negatively affects the SSD endurance and write throughput. This performance loss depends on the drive over-provisioning and the garbage collection method. This paper presents a novel probabilistic model to analytically quantify the impact of over-provisioning on write amplification under a uniformly-distributed random workload and a greedy garbage collection policy. The analysis shows write amplification approximately independent of NAND block size and number of blocks in the SSD. The analysis is verified by full drive simulations.