Résumé
Solid-state drives (SSDs) based on NAND flash memories provide an attractive storage solution as they are faster and less power hungry than traditional hard-disc drives (HDDs). Aggressive storage density improvements in flash memories enabled reductions of the cost per gigabit but also caused reliability degradations. A recent large-scale study revealed that the uncorrectable bit error rates (UBER) in data center SSDs may fall far below the JEDEC standard recommendations. Here, a technique is proposed to improve the tolerated raw bit error rate (RBER) based on the observation that :(a) a small SSD ratio may have a much higher RBER than the rest and (b) the RBER is dominated by the retention error rate. Instead of employing stronger but costly error-correcting codes a statistical approach is used to estimate the remaining retention time, i.e., the reliable data storage time, of flash memory pages. This estimation can be performed each time a memory page is read based on the number of detected retention errors and the elapsed time since data was programmed. The fact that the estimated remaining retention time is smaller than a maximum time interval before the next read operation is an indication that data needs to be refreshed. It is estimated that the tolerated RBER can be increased by more than a decade over a storage period of 3 years if the stored data is verified on a monthly basis and refreshed only if necessary. The proposed technique has the ability to adapt the average time between refresh operations to the actual RBER. This enables performance overhead reductions with factors between 8x and 12x as compared to a conventional scheme with fixed refresh frequency.