On 6/13/11 8:56 PM, "Victor T." <[email protected]> wrote:
>I was making some tests with OpenVSwitch and OpenFlow 1.0 to see what
>happens in the switch (PC) CPU usage when have a CBR flow of data and
>variable packet size for each test.
>
>Is there any work about this?
There are some performance numbers related to transfer rate in a paper we
wrote a couple of years ago:
http://openvswitch.org/papers/hotnets2009.pdf
>Another question, i've noticed that when i start the packet generator
>(iperf), for low packet sizes, the CPU usage of the machine has a peak
>for a few seconds, but then it fades and "stabilizes", when for huge
>packets, the CPU usage is always "low". Is it just a OpenVSwitch
>implementation issue, or there is an explanation to that?
Without knowing your test setup, my initial guess is that your early
packets are hitting userspace before a cached kernel entry is getting set
up. In Open vSwitch, the kernel contains exact-match flow entries that
allow fast hash-based lookups, which are pushed down on-demand from
userspace. The OpenFlow flows are stored in a userspace classifier that
are only consulted on one of these kernel "misses"; the result of which is
blown out into an exact match flow for the kernel.
For future reference, one of the Open vSwitch mailing lists are probably a
better venue for this topic, since this isn't particularly related to
OpenFlow:
http://openvswitch.org/?page_id=17
Hope that helps.
--Justin
_______________________________________________
openflow-discuss mailing list
[email protected]
https://mailman.stanford.edu/mailman/listinfo/openflow-discuss
