Applying fast shallow write to short-lived data in solid state drives

Abstract

In NAND flash memory, a high threshold voltage during incremental step pulse programming speeds up the write operation but reduces the data retention time, and vice versa. Current NAND flash memory uses a low threshold voltage to satisfy the industry standard, which requires data retention of more than a year, and as a result its write operation latency tends to increase as semiconductor process technology progresses. However, actual server workload analysis indicates that much of the data are short-lived and do not require a long retention time. Writing those short-lived data slowly with a low threshold voltage is inefficient. Fast write with a high threshold voltage should be employed for short-lived data and slow write should be employed only for long-lived data. Therefore, this work proposes a method that predicts the lifetime of data based on write request size and selectively applies fast write to short-lived data. The results of evaluations using representative server workloads on an SSD simulator indicate that the proposed method improves an average performance by up to 41.14% compared with the existing method. Further, the increase in total block erasures due to the wrong prediction is limited to 5.60%.