[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openvswitch-discuss
Subject:    [ovs-discuss] Communicate flow-mod info to kernel module
From:       challa () noironetworks ! com (Madhu Challa)
Date:       2014-05-29 1:29:11
Message-ID: CAN_zqfPGrhpzb2ruz23-1et3FxMt63a68RBRHGVEO44j-B=-Og () mail ! gmail ! com
[Download RAW message or body]

Jesse,

I was able to push 130 kpps max and the cpu usage was less than 1%. iperf
was not doing so great in its multithreaded mode, so I had to start
separate instances. No I did not try just updating usage. Let me see if I
can push the packets per second higher than this so the test could be more
realistic.

Thanks.


On Wed, May 28, 2014 at 5:27 PM, Jesse Gross <jesse at nicira.com> wrote:

> On Mon, May 26, 2014 at 5:03 PM, Madhu Challa <challa at noironetworks.com>
> wrote:
> > Jesse,
> >
> > I ran an iperf test (4 client server pairs) with and without the stats
> and
> > there was no noticeable difference in the measured bandwidth. However as
> I
> > increase the number of connections from 1 to 4 I do see the spin lock
> > overhead for this call stack (as a percent of total spin lock overhead on
> > the system) go up from 6% to 16%. I have 2 numa nodes on my system and 32
> > cores. But yes it is not affecting bandwidth.
>
> Do you notice a difference in the total CPU usage with stats on vs.
> off? The significance of the relatively percentage depends on how big
> the total spinlock overhead is, presumably.
>
> Also, did you try it with updating last used but not packets/bytes?
>
> Overall, I'm not too excited about going down this path unless there
> is a really strong performance justification. The potential to be a
> pain in the future seems high...
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://openvswitch.org/pipermail/discuss/attachments/20140528/925583b3/attachment.html>

[prev in list] [next in list] [prev in thread] [next in thread]