KV-FTL: A Novel Key-Value-Based FTL Scheme for Large Scale SSDs

With the emergence of 3D TLC/QLC NAND flash, the capacity of flash-based SSD is growing rapidly, from hundreds of gigabytes to tens/hundreds of terabytes. Accordingly, the Flash Translation Layer (FTL) within such a large SSD is confronted with serious problems that have not ever appeared before. Traditional FTLs either adopt a coarse-grained mapping mechanism, thus facilitating the mapping table being kept in DRAM completely, or adopt a fine-grained mapping mechanism but only keep frequently-accessed mapping information in DRAM depending on the localities of workloads. We argue that both of the above policies are unsuitable for SSDs supplying ultra-large capacity. Firstly, large SSDs introduce so many more mapping entries than ever before that even the coarse-grained mapping mechanism cannot produce a compact enough mapping table to be kept in DRAM completely. Secondly, large SSDs tend to be deployed in data centers to serve IO requests from massive numbers of users under various application backgrounds, where these IO requests exhibit weaker spatial and temporal localities. As a result, the method that keeps frequently-accessed mapping information in DRAM is also impractical for large scale SSDs. In this paper, we propose a novel KV-FTL approach for large scale SSDs, which mostly maps logical addresses to physical addresses via a simple hash function while handling hash collision and out-of-place data update in the traditional manner, i.e., the mapping table. Our KV-FTL is able to accelerate address translation by avoiding loading the mapping table from flash memory to DRAM, thus improving performance, as well as to reduce the write-traffic introduced by the mapping table, thus extending the lifespan of SSDs. To the best of our knowledge, this is the first time the key-value principle has been applied to FTL design. Experimental results show that our KV-FTL extends SSD lifespan by a factor of up to 18.7% with an average of 13.6%; improves read performance ranging from 18.4% to 50.7% with an average of 39% with optimization, and in the case of extremely intensive requests, improves the overall performance for requests with an average of 47%.