Re: CP Proposal - Benchmark + source code

[Paul Eggert]

>From the numbers you published in your email
<http://lists.gnu.org/archive/html/bug-coreutils/2005-08/msg00125.html>
it appears that the big win on your examples with 'cp' comes from
O_DIRECT, and that there's a mild improvement from O_STREAMING.

However, my understanding is that O_STREAMING is a dead end (see
<http://www.kernel.org/pub/linux/kernel/people/rml/O_STREAMING/IN_LIEU>)
and that they have merged posix_fadvise() into Linux 2.6.

Did you try POSIX_FADV_SEQUENTIAL as well?  It seems to me that
POSIX_FADV_NOREUSE | POSIX_FADV_SEQUENTIAL would be a closer POSIX
approximation to what "cp" needs than POSIX_FADV_NOREUSE by itself.

I don't understand why POSIX_FADV_NOREUSE | POSIX_FADV_SEQUENTIAL
would be so much worse than O_DIRECT, if they're properly implemented.
Perhaps you could elaborate?