KHB: A Filesystems reading list

KHB: A Filesystems reading list

Just a thank you for _The Utility of File Names_; a fascinating paper both
for the niftiness of its ideas and the poverty of what they planned to do
with it (a trainer that doesn't run continuously and doesn't collect data
from the VFS interface, a small set of static predictions, using it only
to choose between two filesystems). Most of these restrictions are simply
because their data collection mechanism was so unwieldy: collecting
complete NFS packet dumps is, well, wildly silly. It's probably also the
only viable method on a closed-source kernel.

A possible variation which occurs to me on a whole ten seconds'
consideration (awed silence, yes, I know) would be to turn the trainer
upside down, and have it constantly watch file
access/modification/creation/deletion patterns, cluster them, and try to
classify each cluster by filename. (The latter job seems like one of those
rather rare cases that cries out for a neural net.)

Anyway, just babbling aloud here really, everyone else probably thought of
this years ago and there's some terribly obvious reason why it won't work.