Yepp, Ben has right.
There are a bunch of performance analyses on the Internet that presents
the throughput of software based switches, like OVS, using multiple cores.
If your traffic is really simple (like a packet generator's simple
pattern, i.e., when all header information are the same for all the
packets), and your OVS is also in some kind of simple setup (e.g.,
portforwarding), then you can realize around 15Gbps throughput on one
core (newer Intel Xeon or Core architectures and DPDK driver).
Just try to assign more cores to OVS! In case of kernel space driver,
OVS automatically do this for you, i.e., not merely OVS itself but the
scheduler of your OS, or use user-space drivers like Intel's DPDK, where
you can explicitly assign specific cores to the OVS system (vswitchd per
se).
You can also achieve similar performance with more complex traffic
patterns but the same OVS setup (portforward), when RSS (receive side
scaling) becomes enabled, i.e., when L3 and above header information is
changing in each packet.
On 07/04/2016 11:30 PM, Ben Pfaff wrote:
On Mon, Jul 04, 2016 at 08:52:15PM +0530, Kapil Adhikesavalu wrote:
I believe, at any of point time, only one flow can be processed in a
bridge. We can't do multiprocessing here within a bridge. Correct me if am
wrong.
This is wrong.
_______________________________________________
discuss mailing list
discuss@openvswitch.org
http://openvswitch.org/mailman/listinfo/discuss
