[ovs-discuss] Questions about some performance issues / bandwitch limitation inside openvswitch

Jesse Gross jesse at nicira.com
Mon Jan 31 21:04:21 UTC 2011

On Thu, Jan 27, 2011 at 7:23 AM, Benoit ML <ben42ml at gmail.com> wrote:
> Hi,
> Thoses values were observed by tcpdump (tcpdump -i ethX) on the physical
> host where gre tunnel were configured.

1524 is a very strange value to be seeing.  There are a few possible
situations but assuming the guest sent an MTU sized packet, none of
them result in a packet that is 1524:

With the MTU the same size on the physical network and VM and no path
MTU discovery, IP fragmentation is required and you should get an MTU
sized packet and a smaller one.

If path MTU is enabled, you should see packets of MTU size because the
VM sends smaller packets.  GRE overhead with the default options and
no vlans is 38 bytes.

If the physical network is larger than the VM MTU, you should see a
packet of size VM MTU + 38.

Jumbo frames inside the tunnel is not currently supported and will
result in packets being dropped.

Note that these numbers are based on the Ethernet payload size.
Depending on the options, tcpdump often shows different sizes: either
the size of payload of the deepest protocol it can decode (IP, so 20
bytes less) or full Ethernet frame size (14 bytes more).

> Well I've tested with 9000 and 1500. Evry elements got it (9000 or 1500) :
> - eth inside guest
> - tap interfaces on the host
> - eth on the host (gre tunnel are configured on thoses physical card).
> Can I specify manually the mtu of the gre tunnel ?
> In the documentation, I've seen that i can enable/disable auto-detection but
> not specify it manually.

There are two MTUs that matter here: that of the guest and that of the
host.  Both of these can be adjusted using the normal tools outside of
Open vSwitch.  Path MTU discovery simply tries to inform the guest of
the host's MTU and the tunnel overhead.

More information about the discuss mailing list