Re: [Scst-devel] Performance of block device I/O versus filesystem I/O

[Bart V. A. <bar...@gm...>]

On Sat, Jul 25, 2009 at 7:12 PM, Bart Van Assche
<bar...@gm...> wrote:
> Results:
> * For brd, XFS I/O is typically slower than block device I/O, as
> expected. But XFS I/O can be up to 7% faster than block device I/O,
> which is not expected. This effect is the strongest for the block size
> (4 MB) that is the closest to the CPU cache size (6 MB).

The above result has been obtained on a system with a dual core CPU. I
have repeated this test with the kernel boot parameter maxcpus=1
added. As you can see with maxcpus=1 XFS I/O is systematically slower
than block device I/O on a ram disk (brd.ko). So apparently XFS is
able to use multiple CPU cores more effectively than direct device
block I/O.

Bart.