[ovs-discuss] packet segmentation in ovs

[Jesse Gross]

How are you sending the packets and what exactly do you see with
tcpdump? 1480 is a suspicious length since it is the maximum payload
of an IP packet, not a GRE packet. It sounds like the host is simply
fragmenting the packet before it gets to OVS.

On Fri, Aug 23, 2013 at 5:47 PM, Theodore Elhourani
<telhoura at email.arizona.edu> wrote:
> This is version 1.10.
>
> Setting the MTUs to take into account the additional GRE headers, was
> actually the only way to get packets of   1480<size<1500 to show up on the
> tunnel.
>
>
>
>
>
> On Fri, Aug 23, 2013 at 4:55 PM, Jesse Gross <jesse at nicira.com> wrote:
>>
>> On Thu, Aug 22, 2013 at 5:24 PM, Theodore Elhourani
>> <telhoura at email.arizona.edu> wrote:
>> > If the original packets are of length <= 1400, then they are
>> > successfully
>> > transmitted through the tunnel. Only when a packet is greater than ~1480
>> > it
>> > gets dropped. You are right the ip stack is fragmenting the large
>> > packets,
>> > but because some of them are a notch below 1500 they are passed as is.
>> > For
>> > these packets, adding gre/ip headers would increase their size to >
>> > 1500bytes. Subsequently they are dropped. The reason may be because the
>> > packets are already beyond the ip stack of the host, and no further
>> > manipulation of the packets is done after this point.
>>
>> What version is this? If it is before 1.9 then setting
>> header_cache=false in the tunnel options may help.
>>
>> > Reducing the size of packets, via MTU tweaking, seems to be a common
>> > practice when gre is used. Any ideas on this ?
>>
>> Regardless of the outcome of the above, setting the MTUs to take into
>> account the additional GRE headers will improve performance.
>
>