[ovs-discuss] Implementing 'group type select' algo in the switch

[Ben Pfaff]

It's a hash, so it's probabilistic.  You'll only get a statistically
even distribution if you have a large number of flows.  For two flows,
there is a 50% chance they end up in the same bucket.

On Fri, Jun 12, 2015 at 03:30:37PM +0500, Haleema Sadia wrote:
> My topology is two hosts connected to a switch and I am sending two flows
> from host 1 to host 2. What I want to do is use the select group type with
> two different buckets. But both of my flows are being mapped to one bucket,
> essentially. Shouldn't the select function, which is taking a hash of
> src/dst ports and ips, map each blow to different bucket?
> 
> 
> 
> On Fri, Jun 12, 2015 at 3:20 AM, Ben Pfaff <blp at nicira.com> wrote:
> 
> > The default is a hash (not just an xor) over fields up to L4, but you
> > can configure another method using the extensions implemented by
> > Netronome.
> >
> > On Thu, Jun 11, 2015 at 11:07:56PM +0500, Haleema Sadia wrote:
> > > What is the current ovs version using for select type? As far as I could
> > > understand it is taking a xor of src and destination addresses and tcp
> > > ports? Is it correct?
> > >
> > > On Thursday, June 11, 2015, Ben Pfaff <blp at nicira.com> wrote:
> > >
> > > > On Thu, Jun 11, 2015 at 01:13:22PM +0500, Haleema Sadia wrote:
> > > > > I was wondering if you could help me figuring out where the
> > algorithm for
> > > > > the group type "select" is implemented in ovs switch files? I want to
> > > > > implement a function which takes the hash of the src/dst ip, and
> > src/dst
> > > > > port and selects the corresponding bucket for the packet.
> > > >
> > > > xlate_select_group() in ofproto-dpif-xlate.c
> > > >
> >