[ovs-discuss] ovswitchd : maximum MAC addresses, possibility to increase it ?

[Benoit ML]

Hello,

Thank you for the answers.
So we will compile a new version with '16384', primary for the central
vswitch. Pretty sweat.

About the versions : for the older hypervisor, we are in 1.0 and 1.2.
We have some in 1.4 too. For the central physical vswitch we are in
1.6.1.

We have a plan to upgrade evrything to the last version, but we need
to test the ovsdb-tool to convert evry database to the lastest schema
... and no really time to do it for now ;)

In the architecture we have all hypervisors (~50 hypervs) vsiwtchd
connected, throught gre tunnel (mtu is ok no probleme here), to the
central vswitchd. So by port on the vswitchd we have plenty of MAC
addr.  An arp request take some times.
Modify the  'MAC learning age', only on the central vswitchd, hide the
probleme : less ARP requests for communication between VM already in
the mac table. But well it's just hiding.
Really I don't know if it's related to the "old" version of vswitchd
on some hypervisors, or GRE tunnels, or  mac learning processing ...
If I can help by some data, let me know ;)

Thank you again.

Regards,

2012/10/17 Ben Pfaff <blp at nicira.com>:
> On Wed, Oct 17, 2012 at 10:18:58AM +0200, Benoit ML wrote:
>> We actually using openvswitch in a large virtualised environnement (~
>> 1000 VMs).  It's working verry nicely ! :) We use a centralised
>> physical openvswitch server.
>
> I'm glad to hear that.
>
> Out of curiosity, what version of Open vSwitch are you using?
>
>> By the way, we have some questions and performances issues about the
>> mac-adresses : an ARP request take some times (about 2 ms), and well
>> it's not a big deal, but it's not very clean too. We have increase the
>> 'mac learning age' to 3600 secondes, and it's better.
>
> Do you mean that Open vSwitch is too slow at processing ARP requests?
> I'm not sure how the MAC learning age is related.
>
>> Futher more, I've seen in the documentation (ovswitchd) that the
>> maximum mac addresses per bridge is 2048. I found that pretty low. In
>> the last version is steel the case ?  Can we modify this behaviure,
>> like something to 10 000 ?
>> Because if it's really 2048 we ve reached this limit ...
>>
>> In the source I see a hard coded parameter in "lib/mac-learning.h" :
>> #define MAC_MAX 2048
>> Can we change this parameter without any side effect ?
>
> You can increase it to a larger power of 2, such as 8192 or 16384.
>
>> Another question, what's the algorithm to add a new adresse mac when
>> the table is full ? replace the first (by age) ?
>
> The MAC address least recently seen on the network is replaced.



-- 
--
Benoit