A Prediction-Based Cache Replacement Policy for Flash Storage

Van-Nguyen Pham, Mwasinga Lusungu Josh, Duc Tai Le, Sang-Won Lee, Hyunseung Choo

Abstract: Replacement algorithms in most disk-based operating systems focus on optimizing memory hit counts. For flash storage, such algorithms would incur high replacement costs in terms of time and energy consumption because writing dirty pages to flash memory is costly. Thus, this work proposes an intelligent approach for efficiently balancing the trade-off between cache replacement costs and cache hit rate performance. Our logistic regression-based approach predicts future reference probabilities of pages in the cache to identify candidate pages for eviction. To ascertain our superiority of the proposed system, we conducted rigorous simulations based on online transaction processing workload traces. Simulation results shows that our approach outperforms state-of-the-art methods.