Re: CP Proposal - Benchmark + source code

[Paul Eggert]

Thanks for the further measurements.  It sounds like in the current
version of GNU/Linux, POSIX_FADV_NOREUSE | POSIX_FADV_SEQUENTIAL is
still much slower than O_DIRECT, even with kernel tweaking, so that is
an argument for having a new option for 'cp', 'mv', and 'install'.  It
would be helpful to know, though, why the performance difference is so
great -- can you ask the kernel folks about that?

If we head this direction, one idea is to mimic the iflag and oflag
options that are already in 'dd'.  Thus:

  cp --iflag=direct --oflag=direct SOURCE DEST

would have the effect that you want.  One could also specify other
combinations, e.g., --iflag='direct,noctty'.  Perhaps it would also
make sense to have a combined --ioflag=direct option, for convenience.