[ovs-discuss] Performance of OVS switch

Nicholas Bastin nick.bastin at gmail.com
Sat Mar 21 08:33:03 UTC 2015


>
> > On Mar 20, 2015, at 6:02 AM, Radhakrishnan, Sharmila <
> sharmila.radhakrishnan at intel.com> wrote:
> > I wanted to understand, what is the maximum performance you can get with
> OVS switch.
>

In a high performance environment this is generally directly correlated to
the memory bandwidth available on your system.  The CPU cycles consumed for
forwarding pale in comparison to the limitations incurred by memory copies.
 (This does not apply directly to DPDK, which has slightly different
scaling factors, although memory bandwidth still applies)


> > I have connected 2 VMs to an OVS switch using Tap interface which are on
> the same network. When I sent packets(using PcktGen application) from 1 VM
> to another, I get very low performance.
>
> > For 64byte packet size, the throughput is ~5Mbps, 128B packet size it is
> 9Mbps and 254b packet size it is 16Mbps.
>

This is awful - you should be able to get at least an order of magnitude
better pps performance on even a moderately sized system.  I suspect that
something is wrong with your VM environment.  What hypervisor are you using
(Xen, KVM, etc.) and what do your guests report as their ethernet drivers?
Also what are the specs of your test system?  (What CPU, what motherboard,
how many DIMMs, etc.)

> What is the maximum throughput that can be achieved with an OVS switch.
>

At large packet sizes on quad-channel systems we've been able to run
hundreds of gigabits, although that was pretty specialized (paying
particular attention to memory and CPU architectures, etc.).  Certainly
600k-800k packets per second is not unreasonable without trying too hard to
optimize, so you've got some other problem.

--
Nick
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20150321/84d9aae8/attachment-0002.html>


More information about the discuss mailing list