Selective Flushing of Modified Data for Smartphone Buffer Cache Management

Buffer cache plays an important role to hide the slow performance of smartphone storage systems. However, as buffer cache consists of volatile DRAM, data may be lost upon a sudden power failure situation. To cope with this situation, data in the buffer cache are periodically flushed to storage. Although periodic flushing improves the reliability of smartphone data, it degrades the efficiency of buffer cache greatly as it amplifies write traffic to storage. In this paper, we show that a small size of persistent memory can eliminate the periodic flushing overhead of smartphone buffer cache by making use of some distinct write I/O behaviors of smartphones. Specifically, our write I/O analysis from popular smartphone applications shows that a limited number of hot data account for most file writes in smartphones, whereas a large portion of file data accessed in smartphones are written only once. Based on this observation, we present the selective flushing policy that adds small persistent memory to buffer cache and keeps frequently modified data in persistent memory without flushing to storage. We perform storage flushing only for written-once data, which are prohibited from entering persistent memory for space efficiency. By replaying popular smartphone application I/O traces, we show that the proposed selective flushing policy reduces the storage write I/O by 25.8% on average and up to 37.8%.