Reinforcement Learning-Based Read Performance Throttling to Enhance Lifetime of 3D NAND SSD

With high storage density and low cost per bit, 3D NAND flash has dominated the market of modern solid-state drives (SSDs). Nevertheless, the growing number of stacked layers and the evolving multi-bits-per-cell technology of 3D NAND flash have led to the issue of fast accumulated read disturbance. Read disturbance leads to data bit errors and requires the error correction code (ECC) scheme to remove bit errors to correct data retrieval during read operations. Nevertheless, owing to the process of read retry and reference voltage adjustment of the ECC scheme, the read performance of 3D NAND flash deteriorates as the number of error bits grows. To mitigate the unwanted impact of prolonged read latency and maintain top read performance, modern SSDs actively rewrite stored data to remove error bits when the read latency becomes higher. However, actively removing data bit errors through data rewrites could wear out the SSDs prematurely and ultimately consume the whole lifetime of SSDs. To resolve such concerns, this paper proposes lowering the number of data rewrites through throttling the read performance of 3D NAND flash-based SSDs via reinforcement learning techniques to adaptively meet the performance requirements of different applications. Trace-driven experiments have shown promising results.