实验室博士生周鹏撰写的论文《LazyDM: A Machine Learning-Empowered Hot Data Management to Mitigate Premature Migration in High-Density Hybrid SSDs》被国际权威学术期刊《IEEE Transactions on Computers》录用。论文将于2026年正式发表。
Abstract—Modern high-density flash memory technologies, such as Quad-Level Cell (QLC), commonly employ a hybrid architecture in which a subset of QLC blocks operates in Single-Level Cell (SLC) mode as a cache for the remaining QLC blocks. In this architecture, hot data is initially programmed to the SLC cache and later migrated to the QLC region at block-level granu-larity. However, limited SLC cache capacity and coarse migration granularity can cause hot data that has not yet cooled to be prematurely migrated to the QLC region, significantly degrading SSD performance and lifespan. This phenomenon is referred to as premature migration. To address this issue, this study proposes LazyDM—a machine learning-based hot data management for SLC cache, designed to optimize both performance and reliability in high-density hybrid SSDs. In particular, we develop a machine learning model to predict the duration over which data remains hot (referred to as hotness duration). We devise a hot data allocation scheme that segregates data in the SLC cache based on their hotness durations, ensuring that hot data within each block exhibit comparable cooling times. We further introduce a cooling-aware data migration scheme that efficiently relocates cooled data from the SLC cache to the QLC region, thereby preventing the performance degradation and endurance reduction caused by premature migration. Experimental results show that, compared with state-of-the-art techniques, our LazyDM reduces the hot data migration ratio, average read response time, average write response time, number of garbage collections (GCs), and number of valid page migrations by an average of 55.14%, 15.82%, 23.19%, 21.97%, and 42.83%, respectively. In addition, LazyDM improves the SLC read hit ratio by an average of 19.34%.