A filesystem for next-generation flash disks

Опубликовано 6 сентября 2016, 18:32

Flash memory has found wide application in military, aerospace, and consumer electronics. Although it offers superior read performance and shock insensitivity than conventional disk-based storage and has often been touted as being more energy efficient, it has traditionally not been suitable for high performance write-intensive applications. Recently, however, a number of manufacturers have introduced high performance solid state disks based on flash memory. These devices typically implement the salient features of existing filesystems designed for flash devices, including wear-leveling and log-structured, garbage collected storage management for improved write performance. I will describe work-in-progress at Princeton on a new filesystem, VsFS, that rather than duplicate these features, delegates many of the storage management functions to the underlying device driver and firmware stack which can in turn be tuned independently for a particular hardware platform. We have implemented a prototype for Linux 2.6 that compared to commonly available filesystems such as Ext3 achieves higher performance as measured in I/O transactions per second at lower cost as measured in percentage of host CPU cycles used. Furthermore, by decoupling storage management from the high-level filesystem interface, it simplifies the implementation and, we anticipate, will facilitate a richer interface, including user-level RDMA-like access to individual files. This talk describes joint work with Kai Li at Princeton and David Flynn at FusionIO in Salt Lake City.