>> Unified buffer & VM cache, and some way around the problem of sync
>> writes slowing down the disk.
> The latter is a significant one. When doing installs I typically turn
> off sync metadata writes and suddenly the thing speeds up
> dramatically.  Unfortunately, this is dangerous to do regularly, but
> a proper way to do metadata journaling would make everything ever so
> much more bulletproof.

I think the problem is (almost) inherent to ffs filesystems.  You
_must_ (a) ensure certain ordering constraints on writes or (b) accept
windows during which a crash can mean a seriously trashed filesystem.
Since the ordering relationships involved are quite complex if tracked
in their full glory, (a) more or less means sync metadata writes.

Ever since reading some four-color glossies blabbing about some vendor
filesystem, I've been toying with ideas thus generated.  Perhaps
someday I'll find the time to figure out what's involved and actually
_do_ something... yeah right, when the devil starts a ski resort :-(

					der Mouse

			    mouse@collatz.mcrcim.mcgill.edu