A Probabilistic Data Replacement Strategy for Flash-Based Hybrid Storage System

摘要

Currently, the popularization of flash memory is still limited by its high price and low capacity. Thus, the magnetic disk and flash memory will coexist over a long period of time. How to design an effective flash-hard disk hybrid storage system emerges as a critical issue. Most of the existing works are designed based on traditional cache management approaches by taking the characteristics of flash into consideration. In this paper, we revisit the existing hybrid storage approaches and propose a novel probabilistic data replacement strategy for flash-based hybrid storage system, named HyPro. Different from traditional deterministic approaches, our approach moves the data probabilistically based on the data access pattern. Such a method can statistically achieve a good performance over massive memory operations of modern workloads. We also present the detailed data replacement algorithm and discuss how to determine the probability of data migration in the storage hierarchy consisting of main memory, flash, and hard disk. Extensive experimental results on various hybrid storage systems show that our method can yield better performance and achieve up to 50% improvements against the competitors.