NAND based solid state drives (SSDs) are almost ubiquitously used in safety-critical systems, and recent advances have demonstrated RAID implementations that built on the top of SSDs can effectively enhance the data integrity and reliability. RAID can restore the lost data chunks in case of failures of RAID components (i.e., SSDs in the context), through a process of RAID reconstruction. Specially, online RAID reconstruction allows the RAID system to continue fulfilling user I/O requests during reconstruction. Servicing user I/O requests, however, significantly affects the performance of reconstruction due to contention for the shared SSD bandwidth. This paper proposes a fast online reconstruction method for SSD-based RAID systems, that preferably restores the lost chunks if the replaced SSD device is idle to reduce the reconstruction time, thus minimizing the probability of a second disk failure in the RAID system during reconstruction. Furthermore, it schedules the tasks of restoring data/parity chunks according to the their impacts on other working SSDs in the RAID system, for the purpose of reducing the overall I/O latency. Through a series of experiments based on the selected disk traces of real-world applications, we show that the proposed reconstruction scheme can reduce the reconstruction time by up to 45.6%, and meanwhile cut down the I/O latency by 9.8% on average compared to state-of-the-art methods.

The innovative stacking architecture of 3D NAND flash offers a promising solution to increase the capacity of flash memory and further cut down the per-unit price. Such compact architectures, however, cause varied kinds of errors because of the hardware nature. Specially, retention errors are primary causes of read retries in high-density flash memory, which are induced by charge leakage over time. This paper proposes an empirical mathematical model to estimate the error rate caused by retention errors on the granularity of block, which is the basic program/erase (P/E) unit in 3D NAND flash memory. Specifically, we build the generalized model by considering the factors of layer-to-layer interference, early retention loss, the P/E cycle and the retention time of data of 3D NAND flash memory. Then, we validate the model on four commercial 3D NAND flash products, and the experimental results verify the accuracy of our proposed estimation model. At last, we apply our model in the existing I/O optimization of write scheduling for 3D NAND flash memory, for showing its contributions to better I/O performance.