[Qemu-devel] idea: non-ethernet paravirtual network device

[Qemu-devel] idea: non-ethernet paravirtual network device

[Sassan Panahinejad]

[-- Attachment #1: Type: text/plain, Size: 514 bytes --]

Hi all,

Here's a thought, could we improve network performance by creating a
paravirtual network device which doesn't emulate ethernet? It shouldn't be
too hard to just whack IP packets pretty much directly over a virtio link.
This should improve performance when using a "user" host connection and we
could introduce a tun host connection instead of tap for this setup.

Does anyone have any thoughts on how worthwhile this would be? Would the
performance improvement justify the effort involved?

Thanks
Sassan

[-- Attachment #2: Type: text/html, Size: 569 bytes --]

Re: [Qemu-devel] idea: non-ethernet paravirtual network device

[Stefan Hajnoczi]

On Mon, Jul 25, 2011 at 4:53 PM, Sassan Panahinejad <sassan@sassan.me.uk> wrote:
> Here's a thought, could we improve network performance by creating a
> paravirtual network device which doesn't emulate ethernet? It shouldn't be
> too hard to just whack IP packets pretty much directly over a virtio link.
> This should improve performance when using a "user" host connection and we
> could introduce a tun host connection instead of tap for this setup.
>
> Does anyone have any thoughts on how worthwhile this would be? Would the
> performance improvement justify the effort involved?

My guess is no noticable impact (if you ignore ARP requests).

The Ethernet header is only 14 bytes or so.  We don't calculate any
checksums at that level.  There's probably not much of a win.

Stefan

Re: [Qemu-devel] idea: non-ethernet paravirtual network device

[Hannes Reinecke]

On 07/26/2011 08:04 AM, Stefan Hajnoczi wrote:
> On Mon, Jul 25, 2011 at 4:53 PM, Sassan Panahinejad<sassan@sassan.me.uk>  wrote:
>> Here's a thought, could we improve network performance by creating a
>> paravirtual network device which doesn't emulate ethernet? It shouldn't be
>> too hard to just whack IP packets pretty much directly over a virtio link.
>> This should improve performance when using a "user" host connection and we
>> could introduce a tun host connection instead of tap for this setup.
>>
>> Does anyone have any thoughts on how worthwhile this would be? Would the
>> performance improvement justify the effort involved?
>
> My guess is no noticable impact (if you ignore ARP requests).
>
> The Ethernet header is only 14 bytes or so.  We don't calculate any
> checksums at that level.  There's probably not much of a win.
>
Only lots of pain to be had.
Mainframe used to do this. But abandoned it not, thankfully.

Problem is that you need to patch each and every tool looking at the 
packets to _not_ expecting an Ethernet header.
And patching up DHCP is _not_ trivial.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)

Re: [Qemu-devel] idea: non-ethernet paravirtual network device

[Sassan Panahinejad]

[-- Attachment #1: Type: text/plain, Size: 1626 bytes --]

Ok, thanks guys. I don't think I'll proceed with this idea.

Thanks
Sassan

On 26 July 2011 11:23, Hannes Reinecke <hare@suse.de> wrote:

> On 07/26/2011 08:04 AM, Stefan Hajnoczi wrote:
>
>> On Mon, Jul 25, 2011 at 4:53 PM, Sassan Panahinejad<sassan@sassan.me.**uk<sassan@sassan.me.uk>>
>>  wrote:
>>
>>> Here's a thought, could we improve network performance by creating a
>>> paravirtual network device which doesn't emulate ethernet? It shouldn't
>>> be
>>> too hard to just whack IP packets pretty much directly over a virtio
>>> link.
>>> This should improve performance when using a "user" host connection and
>>> we
>>> could introduce a tun host connection instead of tap for this setup.
>>>
>>> Does anyone have any thoughts on how worthwhile this would be? Would the
>>> performance improvement justify the effort involved?
>>>
>>
>> My guess is no noticable impact (if you ignore ARP requests).
>>
>> The Ethernet header is only 14 bytes or so.  We don't calculate any
>> checksums at that level.  There's probably not much of a win.
>>
>>  Only lots of pain to be had.
> Mainframe used to do this. But abandoned it not, thankfully.
>
> Problem is that you need to patch each and every tool looking at the
> packets to _not_ expecting an Ethernet header.
> And patching up DHCP is _not_ trivial.
>
> Cheers,
>
> Hannes
> --
> Dr. Hannes Reinecke                   zSeries & Storage
> hare@suse.de                          +49 911 74053 688
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
> GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
>
>

[-- Attachment #2: Type: text/html, Size: 2393 bytes --]