Abstract:
Flash based solid state drives (SSDs) have established themselves as a higher-performance alternative to hard disk drives in cloud and mobile environments. Nevertheless, SSDs remain a performance bottleneck of computer systems due to their high I/O access latency. A common approach for improving the access latency is prefetching. Prefetching predicts future block accesses and preloads them into main memory ahead of time. In this paper, we discuss the challenges of prefetching in SSDs, explain why prior approaches fail to achieve high accuracy, and present a neural network based prefetching approach that significantly outperforms the state-of the-art. To achieve high performance, we address the challenges of prefetching in very large sparse address spaces, as well as prefetching in a timely manner by predicting ahead of time. We collect I/O trace files from several real-world applications running on cloud servers and show that our proposed approach consistently outperforms the existing stride prefetchers by up to 800\(\times \) and prior prefetching approaches based on Markov chains by up to 8\(\times \). Furthermore, we propose an address mapping learning technique to demonstrate the applicability of our approach to previously unseen SSD workloads and perform a hyperparameter sensitivity study.
