As Dan pointed out, it would help to know your setup. Is there a controller 
involved? Are you using long-lived flow entries or could there be a 
timeout/refersh issue? What CPU loads are you seeing for the different 
processes etc.

---Guido Appenzeller
[email protected]
M: 650 469 3780
On Wednesday, March 30, 2011 at 2:16 PM, T Ficarra wrote: 
> >> There is definitely something wrong in your setup. The performance of the 
> >> user space switch is limited, but you a system like the one below you 
> >> should get to 50-100 Mb/s or more.
> 
> What is the best approach for troubleshooting the user space switch?
> 
> Thanks,
> Tim
> 
> 
> On Wed, Mar 30, 2011 at 5:03 PM, Guido Appenzeller <[email protected]> 
> wrote:
> > There is definitely something wrong in your setup. The performance of the 
> > user space switch is limited, but you a system like the one below you 
> > should get to 50-100 Mb/s or more.
> > 
> > From an older email on this list:
> > > Here is a (very preliminary) performance comparison between kernel/user
> > > space of OpenFlow v0.9 reference design (not OpenWRT). v0.9 is VERY OLD
> > >  and I don't recommend to use it (there is no much support), but just 
> > > wanted
> > > to give you a rough figure. The kernel space one had roughly x2 higher
> > > forwarding performance than the user space one.
> > > > * Here is a summary of a brief performance test of v0.9:
> > > > - Userspace switch (udatapath)
> > > > Thrgouput: 47.9Mbps
> > > > Flow setup time: 1.074ms (min 0.827ms, max 6.908ms)
> > > > 
> > > > - Kernel switch (ofdatapath.ko)
> > > > Thrgouput: 94.2Mbps (Note: Soekris has only 100Mbps interfaces)
> > > >  Flow setup time: 0.657ms (min 0.622, max 0.738ms)
> > > > 
> > > > (Note: Throughput was measured by TCP iperf. Flow setup time was by
> > > > 10 sec interval ping (50pings)).
> > > 
> > > Note:
> > > - As Yiannis said, v1.0 software reference design no longer has the kernel
> > >  space implementation. OpenVswitch implementation has it.
> > > - The PC we used was Soekris 5501 (http://www.soekris.com/net5501.htm),
> > >  which had only 100Mbps ether interfaces and poor CPU (500 Mhz AMD
> > > Geode CPU, 512 Mbyte DDR-SDRAM, 4x 100Mbps Ethernet ports), but probably
> > > more powerful than most of the OpenWRT boxes.
> > For fast software switching, in most cases the kernel implementation of 
> > OpenVSwitch will be your best bet.
> > ---Guido Appenzeller
> > [email protected]
> > M: 650 469 3780
> > On Wednesday, March 30, 2011 at 6:58 AM, T Ficarra wrote:
> > > I'm running the OpenFlow 1.0.0-rev1 version of the userspace switch on a 
> > > Fedora 7 64 bit OS. The machine is x86_64 with dual multicore intel 
> > > processors. For a network card we're using a 10Gb/s network processor.
> > > 
> > > I'm testing the switch using iperf. Output is on the order of Kb/s, and 
> > > the OpenFlow switch reports excessive speed, message too long, and drops 
> > > packets.
> > > 
> > > Tim
> > >  On Wed, Mar 30, 2011 at 12:39 AM, Dan Talayco 
> > > <[email protected]> wrote:
> > > >  It would be helpful to know a little more about the setup you're 
> > > > using. For example, what versions of hardware, operating system, 
> > > > OpenFlow and what sorts of traffic you're running.
> > > > 
> > > >  Others can comment on the through put that's been seen on any SW 
> > > > implementation. I'm certainly interested to know what configuration can 
> > > > run at 10G.
> > > > 
> > > >  -Dan
> > > > 
> > > >  On Tuesday, March 29, 2011 at 8:09 PM, T Ficarra wrote:
> > > > > Is there any documentation regarding the limits of the userspace 
> > > > > switch?
> > > > >
> > > > > Specifically, I am trying to run some experiments on a 10Gb/s network 
> > > > > and have been experiencing troubles with OpenFlow crashing.
> > > > >
> > > > > Thanks.
> > > > > T Ficarra
> > > > >
> > > > > _______________________________________________
> > > > > openflow-discuss mailing list
> > > > > [email protected]
> > > > > https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
> > > > >
> > > > 
> > > > 
> > > _______________________________________________
> > > openflow-discuss mailing list
> > > [email protected]
> > > https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
> > > 
> > 
> > _______________________________________________
> >  openflow-discuss mailing list
> > [email protected]
> > https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
> > 
> 
> 
_______________________________________________
openflow-discuss mailing list
[email protected]
https://mailman.stanford.edu/mailman/listinfo/openflow-discuss

Reply via email to