Lightweight Data Lifetime Classification using Migration Counts to Improve Performance and Lifetime of Flash-based SSDs

Abstract

This paper presents a policy that effectively classifies data in solid state drives (SSDs) based on their lifetime to reduce write amplification of flash based storage. Although there have been many existing studies for separating data by their lifetime or hotness, they either require hardware interface support or create nontrivial management overhead. In particular, the recently proposed multi-stream technique has a limitation in that data streams must be given by the host. Moreover, data stream separation requires user program implementation or additional computations. In this paper, we propose MiDA (Migration count based Data Age classification), a lightweight and effective data lifetime separation policy. Specifically, it exploits data movement patterns of flash memory blocks so that it can be classified effectively with only negligible overhead. It is implemented at the storage device level, lifting the need for any modification on the host. Using various synthesized and real world database workloads, we evaluate the proposed technique on FEMU, an NVMe SSD emulator. We find that it improves both the performance and lifetime of flash based SSDs.