[ovs-discuss] reg. kernel flow tables and u-space wild card tables

ESWAR RAO eswar7028 at gmail.com
Tue Oct 1 06:05:39 UTC 2013


Hi Justin,

Thanks for the mail.

If my understanding is correct, using the openflow protocol any controller
writes the flows into only u-space flow tables.
So after 1.11 OVS, the flows can be directly added to kernel tables??

Actually I was under impression if flows were to be added directly to
kernel flow tables (pre 1.11 also) then context switching would be less and
fast lookups can be obtained. Please correct me if I am wrong?? Was there
any constraint with this approach???


Thanks
Eswar


On Tue, Oct 1, 2013 at 11:26 AM, Justin Pettit <jpettit at nicira.com> wrote:

>
> On Sep 30, 2013, at 10:01 PM, ESWAR RAO <eswar7028 at gmail.com> wrote:
>
> > Hi All,
> >
> > I am new to openvswitch.
> > Please correct me if my below understanding is wrong:
> >
> > Openvswitch has mainly 2 sets of tables:
> > (1) exact match table/rule table in datapath (kernel)
> > (2) wild card table/flow table in in-memory ( user space)
> >
> > Recently used flows gets added in datapath table which has only limited
> entries and expiry interval is around 5 sec.
> >
> > Through openflow we can add flow entries in only u-space tables which
> has slow lookup compared to kernel exact match tables.
>
> Your description matches pre-1.11 versions of OVS.  Since 1.11, we've
> supported wildcards in the kernel, which, in most configurations, causes
> fewer flows to be sent to userspace.
>
> > Please let me know why we cant add direct flows in kernel flow tables
> since its fast (as per my understanding they can be added through ovs-dpctl
> only) through openflow ?? Since kernel lookup would be fast why we add
> flows to u-space???
>
>
>  The OpenFlow tables and other switch configuration can be quite
> complicated and often don't translate cleanly into kernel flows.
>
> --Justin
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20131001/56920ef7/attachment.html>


More information about the discuss mailing list