> > You may want aufs to do, > > - /au = /ram + /hdd > > - user issues write(2) to /aufs/fileA > > - fileA exists /ram and aufs writes to /ram/fileA. > > - /ram becomes full and ENOSPC occurs. > > - aufs copy-down (or move-down) /ram/fileA to /hdd/fileA and writes the > > originally requested data to /hdd/fileA. > > > > Currently aufs has no such feature. > > Exactly! > > In the meanwhile I've managed to make it work with FUSE-based mhddfs, > but because of the insane FUSE overhead it gives pitiful results.
Thinking about such feature in real world, I wonder how do you think about this sceinario. - user updates fileA ... fileZ, and aufs moves them down to /hdd (as described above). - but several files still exist on /ram. - it means + frequently used and larger files goes to /hdd + less frequently used or smaller files are kept in /ram My point is "why do we keep those less frequently used files in /ram?" I don't think that is what you want. The less frequently used files should go to /hdd, and the more frequently used files should remain on /ram I beleieve (of course as long as the size of /ram accepts them). If (again _if_) aufs decides "how frequently this file is used" by its atime or something, it may be possilbe to move the idle files down to /hdd. Such move-down operation may be triggered by - periodically (in time-line), or - when the free space of /ram reaches the pre-specified watermark. Does it sound much preferable to your behaviour? J. R. Okajima ------------------------------------------------------------------------------ RSA(R) Conference 2012 Save $700 by Nov 18 Register now http://p.sf.net/sfu/rsa-sfdev2dev1
