Frequent Pattern-Based Mapping at Flash Translation Layer of Solid-State Drives

Abstract

In order to reduce the overhead of garbage collection, as well as yield an even erasure distribution of blocks in solid-state drives (SSDs), this paper proposes a mapping approach at flash translation layer (FTL) of the SSD devices. To be specific, it first analyzes certain I/O requests, for mining frequent item sets of logical sector numbers, to direct scheduling the write requests. After that, we map the requests belonging to the same frequent itemset, to the SSD blocks having a relative small erasure count. That is to say, these requests are supposed to be flushed onto the same SSD blocks when they have been ejected from the cache. As a result, the data on the same block are mostly like to be rewritten together in the near future, so that the overhead of garbage collection can be significantly cut down. Furthermore, it is able to yield a uniform erasure distribution among all blocks of SSDs even without wear-leveling optimization, to purposely extend the lifetime of the SSDs. Through a series of simulation experiments with a number of write-intensive disk traces of real-world applications, we conclude that this newly proposed mechanism does work, as expected.