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

[Ming Lei]

On Thu, Jun 26, 2014 at 11:57 PM, Paolo Bonzini <address@hidden> wrote:
> This is indeed a difference between the ioq-based and block-based backends.
> ioq could submit more than one request with the same io_submit system call.
>

Yes, I have been thinking that is advantage of qemu virtio dataplane, but
it isn't any longer, and qemu dataplane has not any performance
advantage compared with lkvm now.

> 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.

>
> Note that some fallout of the conversion was expected.  Dataplane told us
> experimentally what level of performance could be reached, but was a dead
> end in terms of functionality.  Now Stefan added a whole lot of
> functionality to dataplane (accounting, throttling, file formats and
> protocols, thread-pool based I/O, etc.) and we need to bring back any
> performance we lost in the process.

These features are very good, but looks the conversion is a bit early, :-(

>
> Out of curiosity, how does aio=threads (after the patch) fare in comparison
> to aio=native (before the patch)?

Looks no obvious difference.

Thanks,
-- 
Ming Lei