[ovs-discuss] Openvswitch performance problem

Jose Luis Franco Arza jlfarza15 at gmail.com
Sat May 28 16:50:25 UTC 2011


I have been doing a small research about how many flows per second could
handle the NOX controller and the surprise was that the bottleneck was in
the OpenVSwitch.
When trying to install 20000 flows (modifying the UDP port number per packet
sent) after reaching the rate of 2000 packets per second then the
ovs-openflowd wasn't able to handel all the income packets and it started to
appear packet lost.
Curiously, if the NOX controller was connected to the OvS then all the
packets were forwarded correctly while the NOX was installing the flows in
the flow table. But when all the flows were already installed then it
started to lose packets again.
I checked the CPU load for the ovs-openflowd process and exactly, while the
NOX controller was installing the flows the CPU load was normal, but when
all of them were installed then it was working at 100% or even more (2
cores).
So, I decided to try withot the existance of the NOX controller installing a
unique flow that would forward every packet in through the output port using
the wildcards. And the results said that the openvswitch is not able to
handle more than 2000 flows/sec.
I would like to ask you if is there any limitation in the openvswitch for
the rate or if some of you know what is the reason why it spends so much
processing to forward the packets when the port is changing for each
different packet.
PD:Attached to the email is a document explaining the experiments deployed
and the results of them.

Thank you very much for your help,

Jose
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20110528/b8b445bb/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cpu_ovs.pdf
Type: application/pdf
Size: 313712 bytes
Desc: not available
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20110528/b8b445bb/attachment-0002.pdf>


More information about the discuss mailing list