[ovs-dev] Megaflows Question

[Jarno Rajahalme]

> On Sep 8, 2015, at 9:38 AM, Ben Pfaff <blp at nicira.com> wrote:
> 
> On Tue, Sep 08, 2015 at 09:41:51AM +0300, Gal Sagie wrote:
>> for example: priority 100 match 10.0.0.0/24  go to controller
>>                   priority 100 match 10.0.0.1 go to next table
> 
> That's a bad idea: these flows overlap so they should have different
> priorities.

More specifically, assuming you want longest prefix match, you should give higher priorities to longer prefixes. One way to do that is to add the prefix length to the priority, i.e., for 10.0.0.0/24 use priority 124 and for 10.0.0.1 use priority 132.

  Jarno

> 
>> I am sending a generated packet using ovs-appctl with destination IP that
>> doesnt match
>> any subnet or specific IP in table 2 :
>> 
>> ovs-appctl ofproto/trace br-int
>> in_port=2,dl_src=00:00:00:00:00:01,dl_dst=fa:16:3e:e2:16:63,eth_type=0x800,nw_dst=8.8.8.8,nw_src=192.1.1.1
>> -generate
>> 
>> The installed megaflow i see is this:
>> Megaflow: recirc_id=0,ip,in_port=2,dl_dst=fa:16:3e:e2:16:63,nw_dst=
>> 0.0.0.0/2.0.0.0,nw_frag=no
>> 
>> The question is how does the nw_dst is calculated in this megaflow ?
>> is it finding a mask that doesnt fit any IP or subnet in table 2 and
>> calculate it? (it seems
>> like a hard task, especially with combined IP's and subnets matchers).
> 
> Yes, that's what it's doing, see "Prefix Tracking" in classifier.h as
> well as the more formal description in our NSDI 2015 paper at
> http://openvswitch.org/support/papers/nsdi2015.pdf.  It's not all that
> expensive, don't worry about it.
> 
>> If i add a learning flow to this table with nw_dst of 8.8.8.8 (for example
>> do learning for every miss in table 2) is it going to be more efficient
>> performance wise? (because the megaflow would then be able to have nw_dst
>> of 8.8.8.8)
> 
> Adding a flow for every actual nw_dst is going to be much much more
> expensive unless there are very few actual nw_dst.
> _______________________________________________
> dev mailing list
> dev at openvswitch.org
> http://openvswitch.org/mailman/listinfo/dev