Just a follow up, in case others are interested in this.
We ran into this problem in another experiment. We did a little bit of
digging and indeed the HP switches at least have a rate limiter for the
software path that is set to 100 pkts/sec by default.
We did modify it and set it up to 1000pkts/sec while monitoring the CPU.
Except from temporary spikes of up to 24% usage, the CPU usage was
pretty low. The flow modified only a couple of hdr fields so I am not
sure how more
complicated actions would affect that.
--niky
Ben Pfaff wrote:
Yes, the PC running OVS would have to be used as a switch. If one of
your requirements is that you need more than the number of Ethernet
ports you can conveniently put in a PC chassis (which is often 6 to 10
ports), then it's not a good way to go.
Presumably the hardware switch's performance drops because the ASIC
can't do what you are asking and every packet has to be sent to the
switch CPU. Switch CPUs are usually quite slow and often the channel
from ASIC to CPU is ratelimited too. With OVS, every packet already
goes up to the CPU (that's a NIC's job after all) and there's very
little incremental cost to modifying a few headers.
On Sun, Jun 26, 2011 at 03:15:48PM -0400, Aaron Rosen wrote:
But in order to use OVS in this manor I would need a box running OVS
that everyone first sent their traffic to first right? Is the reason
why OVS can perform a factor of 1000 faster is because the
implementation is better or because the hardware of a commodity PC is
better for this than that of the OF switch that is doing the
modifying? Just curious.
On Sun, Jun 26, 2011 at 3:00 PM, Ben Pfaff <b...@nicira.com> wrote:
On Sun, Jun 26, 2011 at 02:54:13PM -0400, Aaron Rosen wrote:
Thanks for your reply. I just tested this on the same HP just
rewriting the DL_DST and the performance is the same there too so it
must be done in software completely on these switches :(
You mentioned a "high" rate of 683 Mbit/sec. If that's a good rate for
you, you don't need switching ASIC for that. Open vSwitch can handle
several gigabits per second on commodity PC server hardware. And its
performance won't drop by a factor of 1000 when you start modifying
headers.
_______________________________________________
openflow-discuss mailing list
openflow-discuss@lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
_______________________________________________
openflow-discuss mailing list
openflow-discuss@lists.stanford.edu
https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
