S-FTL: An efficient address translation for flash memory by exploiting spatial locality

The solid-state disk (SSD) is becoming increasingly popular, especially among users whose workloads exhibit substantial random access patterns. As SSD competes with the hard disk, whose per-GB cost keeps dramatically falling, the SSD must retain its performance advantages even with low-cost configurations, such as those with a small built-in DRAM cache for mapping table and using MLC NAND. To this end, we need to make the limited cache space efficiently used to support fast logical-to-physical address translation in the flash translation layer (FTL) with minimal access of flash memory and minimal merge operations. Existing schemes usually require a large number of overhead accesses, either for accessing uncached entries of the mapping table or for the merge operation, and achieve suboptimal performance when the cache space is limited. In this paper we take into account spatial locality exhibited in the workloads to obtain a highly efficient FTL even with a relatively small cache, named as S-FTL. Specifically, we identify three access patterns related to spatial locality, including sequential writes, clustered access, and sparse writes. Accordingly we propose designs to take advantage of these patterns to reduce mapping table size, increase hit ratio for in-cache address translation, and minimize expensive writes to flash memory. We have conducted extensive trace-driven simulations to evaluate S-FTL and compared it with other state-of-the-art FTL schemes. Our experiments show that S-FTL can reduce accesses to the flash for address translation by up to 70% and reduce response time of SSD by up to 25%, compared with the state-of-the-art FTL strategies such as FAST and DFTL.