Re: [Xen-users] Xen Performance

[Grant McWilliams <grantmasterflash@xxxxxxxxx>]

On Fri, Oct 16, 2009 at 5:43 AM, Dustin Henning <Dustin.Henning@xxxxxxxxxxx> wrote:

<snip> DomU to DomU performance is 1/3 to 2/3 that of DomU to Dom0 even though that the DomU to DomU traffic is probably traversing the same path as DomU to Dom0. <snip>

Cheers!
Grant McWilliams
______________________________________

Seeing as how each domU nic exists in dom0, as does the bridge, I would argue that traffic between domU and domU take 50% more steps than traffic between domU and dom0.  Looking at it the other way, that would be 33% less steps between domU and dom0 than domU and domU.  Here is why:

Between domU A and domU B, the traffic has to traverse domU A's vif in dom0, the bridge in dom0, and domU B's vif in dom0.  That is three virtual devices.

Between domU A and dom0, the traffic only has to traverse domU A's vif and the bridge, it has then arrived in dom0, that is only two virtual devices.

I could be wrong, and would love to be given a more detailed and technical answer as to how I am if that is the case.  Regarding the performance difference, my opinion is that we could always use more performance in all aspects, period.  Sure, eventually costs are prohibiting and we have to settle, but more performance wouldn't hurt no matter how many people think performance is good enough, and only seeing notably better performance can convince some people that any given performance isn't good enough.

Dustin

Dustin,
   This is how it works... :-) This entire thread exists somewhere else since I went through the exact same process the first time I was doing my testing. I understand how Xen works and it's unfortunate that this limitation will never be replaced. Maybe it will but by then everyone will have migrated to KVM which doens't seem to suffer from it.  It's very inefficient to pass all traffic to Dom0 then pass it back to another domU. It's like you always have a router to go through even if the DomUs are on the same physical network.

It also seems that it's the return trip that's slow (even without the broken pipe) and my testing before (and the resulting message thread before) showed that data going to the Dom0 was fast but data back wasn't. If we throw enough hardware at it we get wire speed or we assign a nic to each DomU but then we can only get wire speed.


Grant McWilliams

Some people, when confronted with a problem, think "I know, I'll use Windows."
Now they have two problems.

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users