Balance Performance and Cost: A Cold/Hot Data Classification Algorithm for NVM-SSD Hybrid Storage

Modern data-intensive applications, exemplified by IoT and AI systems, are driving exponential growth in storage demands, which demands efficient storage hierarchies that balance performance and cost. Traditional DRAM-SSD architectures face limitations: DRAM is volatile and expensive, while SSD suffers from latency and write endurance issues. Against this backdrop, Non-Volatile Memory (NVM) has emerged as a revolutionary technology, bridging the performance gap between memory and storage. With near-DRAM latency and byte-addressability, NVM offers an attractive solution for data-intensive workloads. However, despite its excellence in handling hot data, NVM remains an expensive resource compared to SSD. SSD, with their mature technology, high capacity, and continuously declining cost per GB, continue to play an irreplaceable role in storing cold data. We propose a cold data management solution for NVM-optimized databases. By logging NVM accesses and monitoring hit rates, our data migration strategy uses a Boltzmann distribution probability model to distinguish between hot and cold data and lazily migrate them to the corresponding storage media. The experimental results validate the accuracy and the low overhead.