Re: [Qemu-devel] [regression] dataplane: throughout -40% by commit 580b6b2aa2

[Kevin Wolf]

Am 27.06.2014 um 06:59 hat Paolo Bonzini geschrieben:
> Il 27/06/2014 03:15, Ming Lei ha scritto:
> >On Thu, Jun 26, 2014 at 11:57 PM, Paolo Bonzini <address@hidden> wrote:
> >>We can implement (advisory) calls like bdrv_plug/bdrv_unplug in order to
> >>restore the previous levels of performance.
> >
> >Yes, that is also what I am thinking, or interfaces like bdrv_queue_io()
> >and bdrv_submit_io(), which may match with aio interfaces.
> 
> Would you like to try preparing a patch?

Note that there is already an interface in block.c that takes multiple
requests at once, bdrv_aio_multiwrite(). It is currently used by
virtio-blk, even though not in dataplane mode. It also submits
individual requests to the block drivers currently, so effectively it
doesn't make a difference, just the problem occurs in the block layer
instead of the device.

We should either improve bdrv_aio_multiwrite() to submit the requests in
a batch to the block drivers, add a bdrv_aio_multiwrite() and use it for
dataplane as well (possibly with a flag for disabling the request merging
if we want to keep the current behaviour for dataplane); or, if we
consider it a bad interface, replace it altogether with the new thing
even for normal virtio-blk.

If this makes a difference for dataplane, it probably makes a difference
for all block devices.

Kevin