Re: [Xen-users] slow network with gplpv drivers in vlan setup

[Matthieu Patou <mat+Informatique.xen@xxxxxxxxx>]

On 09/02/2010 13:03, James Harper wrote:
> > > Large Send is probably what is causing you problems. Some network
> > >        
> cards
>    
> > > support large send offload only for untagged packets.
> > >        
> > Well that's a bit strange because on 2 identical dom0 servers
> >      
> configured
>    
> > the same way I had this behavior only on one dom0 (the really loaded
> > server).
> >      
> How identical is identical? Are you able to determine the exact chipset
> and firmware version of the network adapters? Even if you bought the two
> servers from the same supplier at exactly the same time, there is still
> a chance that there is some hardware difference, most likely firmware.
>    
Well identical like the same server at the same moment, for the firmware 
I don't know really how to check it for the nic (any tip ?).
>    
> > > Linux doesn't seem
> > > to know this though so the result is just that it doesn't work. Try
> > > turning off large send and see what happens.
> > >
> > >
> > >        
> > Ok so you suggest to change the value of 61440 to something smaller
> >      
> (ie.
>    
> > 8192)
> >      
> It may be worth a go, but I suspect that the difference would be in
> turning it off or on, not the size. Unless the underlying chipset has
> some limitation in size... I hadn't considered that.
>
>    
> >      
> > > Turning off scatter gather will (almost completely) disable large
> > >        
> send
>    
> > > also, because windows is then limited to a total packet length of
> > >        
> 4096
>    
> > > bytes which can't be broken up unless the MSS is really small.
> > >
> > >        
> > The question that stay is what is drawback of not having  large send ?
> >
> >      
> Large send means that the network card will accept TCP packets well in
> excess of the actual MTU, up to about 60K. The network card computes
> checksum, seq, etc for you. So if you want to send a lot of TCP data
> it's the difference between windows giving one 60K packet to the network
> card vs 40 packets.
>
> It gets even better when you are talking about virtual machines because
> a Linux Dom0 can keep the packet 'large' as long as all the things it
> has to pass through can handle it, whether that's from the DomU to Dom0,
> DomU through the bridge to another DomU, or DomU through the bridge to
> the physical network card.
>
> In the testing I've done it's been the difference between 2GBits/second
> iperf throughput and 3-4GBits/second. That's probably not representative
> of real-world workloads though.
>
> So in turning it off you do lose out on performance, but only if it
> worked in the first place, which it doesn't for you.
>
> If you could figure out exactly what is different between you're 2
> Dom0's I'd be grateful. I keep getting these reports of LSO causing
> problems for some people and have never been able to properly figure out
> exactly why...
>    
As I said in some other mails I have also the same behavior on a test 
server that is in fact a Dell workstation with a broadcom gigabit nic.
Once I know how to search for the firmware of the nic I'll try to dig on it.

Matthieu.

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