Optimal SSD Management with Predictions

Flash-based solid state drives (SSDs) have emerged as a leading storage technology, but their unique constraints pose fundamental management challenges. Unlike hard-disk drives, SSDs do not allow in-place updates: data must be written to a clean location within a block, and only entire blocks can be cleaned. Moreover, when a block is cleaned, all valid data it stores must first be rewritten elsewhere, resulting in additional writes that harm both performance and device longevity. In this work, we present one of the few worst-case analyses of algorithms for minimizing rewrites in SSDs. Our model falls within the framework of learning-augmented algorithms: algorithms that receive advice from a prediction oracle and their performance is analyzed as a function of the oracle's error. We introduce a novel algorithm for this model, and prove that it achieves optimal worst-case guarantees. Experiments on a range of SSD workloads confirm the practical advantages of the proposed algorithm over existing SSD management schemes.