Open Stack

On Wed, 2014-01-29 at 10:35 -0500, David Magda wrote:
> On Wed, January 29, 2014 10:25, Alvise Dorigo wrote:
> [...]
> > So, it seems that the DHCP client's request correctly originates from the
> > VM, and is correctly trasmitted to the network node by mean of the data
> > network the two br-int interfaces are connected to; but for some reason
> > the dhcp-agent doesn't receive anything, as proved by the null traffic on
> > its tap.
> >
> > Any idea about what and how I could more deeply investigate ?
> 
> On your network node, check the NIC settings for various offloading
> "optimizations":
> 
>     $ ethtool -k eth0 | grep offload
>     tcp-segmentation-offload: off
>     udp-fragmentation-offload: off [fixed]
>     generic-segmentation-offload: off
>     generic-receive-offload: off
>     large-receive-offload: off [fixed]
>     rx-vlan-offload: on
>     tx-vlan-offload: on
> 
> We recently solved this problem at $WORK by disabling GRO on our network
> node.
> 
> The other issue we've come across that may be related is MTUs: since Open
> vSwitch 1.9+ got rid of Path MTU Discovery, you may need to adjust your
> MTUs:
> 
>   http://techbackground.blogspot.ca/2013/06/path-mtu-discovery-and-gre.html
> 
> There are two ways to fix this: if your network infrastructure has jumbo
> frames (1500+) enabled, change the hardware NICs to have an MTU of 1546
> (or larger). If you do not have jumbo frames enabled, configure the guests
> to have an MTU of 1454 or smaller.
> 
> One or both of these things (GRO and MTU) will probably be needed for a
> properly functioning virtual network.

Anne, it would be great to add David's findings/suggestions above to the
operator's handbook. Also, Darragh's blog post contains some excellent
pointers around determining PMTU for network segments and how GRE
overlay networks are affected by adjustments in that.

Best,
-jay