Superblock management is widely used in commercial products of flash memory, in which multiple blocks from different parallel units are organized as a superblock, to provide high-level access parallelism. However, excessive parallelism inadvertently amplifies the negative effects of read reclaim (RR) due to the larger scale of data migrations and space reclamation. This is because a superblock-granularity RR operation must be triggered when any block inside a superblock reaches the RR threshold, even though other blocks are far from the threshold. To improve the RR efficiency for superblock management, this paper proposes a shuffling-based RR scheme for unifying read accesses across all component blocks inside the superblock without any additional space costs. Specifically, it shuffles the blocks of superblock with the granularity of card that consists of multiple data pages in the RR process and ensures no impact on access parallelism after shuffling. As a result, all component blocks may serve a similar number of read requests and approach the RR threshold synchronously. Trace-driven simulation experiments demonstrate that our proposed method can reduce the RR operations and read latency by 14.5 % and 12.9 %, compared with conventional superblock-granularity RR schemes in flash memory, while preserving access parallelism.
Liao et al. (Fri,) studied this question.