[ovs-discuss] How to set the best setting to test packet processing performance?

Ben Pfaff blp at nicira.com
Sat Jul 18 23:42:29 UTC 2015


As the paper says, we used one server with two 10-Gbps NICs.

Please read the paper:
        http://openvswitch.org/support/papers/nsdi2015.pdf

On Sat, Jul 18, 2015 at 04:35:02PM -0700, Hao Wu wrote:
> Hi Ben,
> 
>    I tried netperf. I run 40 netperf instances in 40 different hosts
> connecting to one OVS. For every two host, I set one as the server and the
> other one as the client that sends out packets. Thus, I have 20 pairs of
> C/S. In this scenario, I get 1.27Mpps. I also tried to only hold one server
> and let other 39 host send packets to it in parallel, but the throughput is
> less than the former case. Which is the one you take in your setting? I
> even add more hosts, say, 100 hosts. But Mininet crashes when OVS adds the
> 100 ports connecting to the 100 hosts.....
> 
> +++++++++++++++++
> Best,
> Hao
> 
> On Thu, Jul 16, 2015 at 1:14 PM, Ben Pfaff <blp at nicira.com> wrote:
> 
> > On Thu, Jul 16, 2015 at 01:12:49PM -0700, Joe Stringer wrote:
> > > On 16 July 2015 at 13:08, Ben Pfaff <blp at nicira.com> wrote:
> > > > On Thu, Jul 16, 2015 at 11:22:29AM -0700, Hao Wu wrote:
> > > >>    Yes, you are right. I find the bottleneck is tcprelay which only
> > > >> generates packets at 750Kpps. But even when I add more hosts to send
> > > >> packets in parallel, I can't get a higher generation rate. E.g., if I
> > use
> > > >> only one host, tcpreplay sends packets at 300Kpps, while if I use 6
> > hosts,
> > > >> each tcpreplay sends packets at around 130Kpps and the total rate is
> > still
> > > >> 750Kpps. How do you get a higher generation rate in your experiment?
> > Thanks.
> > > >
> > > > I think we used netperf.  I've never use tcpreplay so I don't have any
> > > > hints.
> > >
> > > Even with netperf, you need to run several threads and depending on
> > > the test you may exhaust all CPU running netperf to generate the
> > > traffic before you start hitting OVS performance limits.
> >
> > Right, the OVS paper mentions that we ran 400 netperf instances in
> > parallel.
> >



More information about the discuss mailing list