[ovs-discuss] Slow serial console seems to drastically affect OVS performance

Ben Pfaff blp at ovn.org
Fri Nov 25 20:16:56 UTC 2016

On Fri, Nov 25, 2016 at 04:23:13PM +0200, Jarkko Oranen wrote:
> I tried sending this to bugs at openvswitch.org first but it bounced. Is that
> address no longer available?

It should still work, but I see the bounces also.  I filed a ticket with
the Linux Foundation to get it looked at.

> I posted to ovs-discuss earlier about performance issues with OVS on CentOS 7.
> See <https://mail.openvswitch.org/pipermail/ovs-discuss/2016-November/043003.html>
> It seems we've found the culprit, and it's quite unexpected. Due to some
> broadcast in our networks, OVS is at times receiving jumbo packets which cause
> the following to appear in dmesg:
> [1991029.408139] openvswitch: qvo41ed92a6-50: dropped over-mtu packet: 8982 > 1500
> [1991029.765021] openvswitch: qvofc43bfd2-9d: dropped over-mtu packet: 8982 > 1500
> [1991031.021320] openvswitch: qvo41ed92a6-50: dropped over-mtu packet: 8982 > 1500
> On a host where the serial console speed is set to 115200, these log messages
> print out fast enough that no significant harm is done, but on the hosts where
> we had these issues, our serial console was misconfigured and very slow. This
> seems to have affected OVS enough to cause network latency.
> After fixing the console settings, we're no longer experiencing the heavy
> latencies, probably because it's now logging fast enough not to trigger bad
> behaviour...
> Is there perhaps a way to avoid this issue altogether?

I think that serial logging from a softirq, which is where OVS handles
packets, is always going to happen synchronously, which means that it's
going to be as slow as the serial connection.  Here's an example of a
similar issue:

Maybe someone else has a good idea, though; I'm no expert here.

More information about the discuss mailing list