Latest Research by TSYS School's Yi Zhou Improves Performance and Reliability in High-Density Hybrid Solid-State Drives

Modern high-density flash memory technologies, such as Quad-Level Cell, commonly employ a hybrid architecture in which a subset of Quad-Level Cell blocks operates in Single-Level Cell mode as a cache for the remaining Quad-Level Cell blocks. In this architecture, hot data, or active, business-critical information that is accessed frequently and requires rapid, low-latency retrieval, are initially programmed to the Single-Level Cell cache and later migrated to the Quad-Level Cell region at block-level granularity. However, limited Single-Level Cell cache capacity and coarse migration granularity, which refers to transferring or moving large blocks of data, objects, or units at one time rather than small, individual pieces, can cause hot data that has not yet cooled to be prematurely migrated to the Quad-Level Cell region, significantly degrading solid-state drive performance and lifespan. This phenomenon is referred to as premature migration. To address this issue, a new study by TSYS School computer scientist Yi Zhou and researchers from Jinan University and Auburn University proposes LazyDM — a machine learning-based hot data management for Single-Level Cell cache, designed to optimize both performance and reliability in high-density hybrid solid-state drives. In particular, Zhou and his colleagues develop a machine learning model to predict the duration over which data remains hot (referred to as hotness duration). The researchers devise a hot data allocation scheme that segregates data in the Single-Level Cell cache based on their hotness durations, ensuring that hot data within each block exhibit comparable cooling times. Their study, which is set to appear in an upcoming issue of IEEE Transactions on Computers, further introduces a cooling-aware data migration scheme that efficiently relocates cooled data from the Single-Level Cell cache to the Quad-Level Cell region, thereby preventing the performance degradation and endurance reduction caused by premature migration. Experimental results show that, compared with state-of-the-art techniques, the researchers' LazyDM reduces the hot data migration ratio, average read response time, average write response time, number of garbage collections, and number of valid page migrations by an average of 55.1%, 15.8%, 23.2%, 22.0%, and 42.8%, respectively. In addition, LazyDM improves the SLC read hit ratio by an average of 19.3%.