Abstract |
Flaf ("First Load All Files") is a new filesystem that tries to achieve modern features like crash robustness and high metadata performance but small code size by adapting to the ever-widening gap between sequential and seek speeds of modern disks.
Contemporary filesystems like SGI xfs and Veritas VxFS that provide crash robustness by way of journalling, and good metadata performance using indexed directories, are complicated and massive in terms of code size. Much of this complexity deals with the synchronization of heavy state between memory and disk.
It turns out that because the relationship between (memory size, sequential disk bandwidth, number of files stored) very often tends to stay within certain bounds, it is feasible to sidestep the complexity of the likes of xfs by having virtually no interdependencies on-disk (no directory lists, no block bitmaps), but instead building the structure at mount time.
The basis of flaf was the idea of spending "5-10 seconds at mount time and 5-10 percent of physical memory" in return for small code size in combination with modern features like crash robustness and good metadata performance.
In flaf, all metadata resides in a single disk block per file and everything else follows from that.
|
