[ovs-discuss] Flow Expiration returning zeros for byte/packet counts

[Jesse Gross]

David Erickson wrote:
>>
>> 1. If you query the flow stats before the expiration message, do you 
>> get the right results?  You mentioned that this was true before but 
>> these flows are less than your 10 second polling interval.
>
> The values still look wrong from the stats message, I just did a test 
> where I bumped up the polling rate so I would send a stats message 
> while the flow was idle but before timeout and got the following:
>
> Nov 12 16:14:01 epic nox: 00087|Flow_tracker:DBG Posting an update 
> event- dpid: 003048b06117 duration: 16 pkt count: 58467 byte count 
> 254630713
>
> Nov 12 16:14:02 epic nox: 00101|Flow_tracker:DBG Flow expiration from 
> dpid: 003048b06117 duration: 16 pkts: 58467 bytes: 254630713 flow: 
> port0002:vlanffff mac22:db:ed:04:cb:a4->7a:4a:0f:e9:1b:bd proto0800 
> ip10.79.2.13->10.79.2.12 port80->54241

Can you also send me the output of ovs-dpctl dump-flows while running 
this test?  Hopefully it isn't too hard to hit the time window.  So:
ovs-dpctl dump-flows <bridge>

>> 2. Does the problem also occur if you use wildcarded rules instead of 
>> exact match?
>
> It actually gets worse, I wildcarded vlan and the expiration messages 
> now have 0's, but the stats message is similar to the above listed 
> message.

It looks like what is happening is that two expiration messages are sent 
out - one for the actual wildcarded flow with the correct stats and one 
or more for the exact match flows from actual traffic with zeroed 
stats.  I sent out patch for review which suppresses the exact match 
flows from wildcarded entries.

>
>> 3. When do you insert the flows?  Before any packets are sent?  In 
>> response to a packet in?
>
> In response to a packet in.

Is it possible that your controller is adding the same flow more than 
once?  Doing so would zero out all that stats on the previously add rule.

>>
>> 5. Do you see similar results with NetFlow?  You can use Wireshark as 
>> a simple NetFlow collector if you don't have one - just tell it to 
>> decode the traffic as cflow.
>
> Can you give me more information on how to do this? IE are netflow 
> packets automatically broadcast out?  IE if I start up tcpdump on 
> xen's dom0 which interface should I listen to? Do I need to explicitly 
> tell OVS somehow to pump out netflow messages?
>

The easiest thing to do is set the key 
netflow.<bridge>.host=<host>:<port> to some machine running Wireshark 
and then tell it to decode that destination port as cflow.  You could 
also point it to any random place and capture from dom0.  The packets 
will go out dom0's IP stack, so whatever interface that is on the 
appropriate network (probably xenbr0) is where you want to capture 
from.  I don't think that tcpdump itself knows how to decode NetFlow though.

Thanks,
Jesse